ch4 - Weste04r4.fm Page 67 Monday, January 5, 2004 1:24 AM...

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: Weste04r4.fm Page 67 Monday, January 5, 2004 1:24 AM 4 4.1 Introduction In Chapter 1 we learned how to make chips that work. Now we move on to making chips that work well, where “well” can be defined as fast, low in power, inexpensive to manufacture, reliable, etc. Before we can choose which design alternative is best, we must develop ways to estimate the goodness of each option, especially with regard to speed and power consumption. The most obvious way to characterize a circuit is through simulation, and that will be the topic of Chapter 5. Unfortunately, simulations only inform us how a particular circuit behaves, not how to change the circuit to make it better. Moreover, if we don’t know approximately what the result of the simulation should be, we are unlikely to catch bugs in our simulation model. Mediocre engineers rely predominantly on computer tools, but outstanding engineers develop their physical intuition to rapidly estimate the behavior of circuits. In this chapter we are primarily concerned with the development of simple models that will assist us in the understanding of system performance. In a modern process, interconnect performance can be as important as or even more important than transistor performance. The issues to be considered in this chapter are Delay estimation in CMOS gates Power dissipation of CMOS logic Interconnect delay and signal integrity Design margining Reliability Effects of scaling 67 Weste04r4.fm Page 68 Monday, January 5, 2004 1:24 AM 68 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION 4.2 Delay Estimation In most designs there will be many logic paths that do not require any conscious effort when it comes to speed. However, usually there will be a number of paths, called the critical paths, that require attention to timing details. These can be recognized by experience or timing simulation, but most designers use a timing analyzer, which is a design tool that automatically finds the slowest paths in a logic design (see Section 8.4.1.5). The critical paths can be affected at four main levels: The architectural/microarchitectural level The logic level The circuit level The layout level The most leverage is achieved with a good microarchitecture. This requires a broad knowledge of both the algorithms that implement the function and the technology being targeted, such as how many gate delays fit in a clock cycle, how fast addition occurs, how fast memories are accessed, and how long signals take to propagate along a wire. Tradeoffs at the microarchitectural level include the number of pipeline stages, the number of execution units, and the size of memories. The next level of timing optimization comes at the logic level. Tradeoffs include types of functional blocks (e.g., ripple carry vs. lookahead adders), the number of stages of gates in the cycle, and the fan-in and fan-out of the gates. The transformation from function to gates and registers can be done by experience, by experimentation, or increasingly by logic synthesis. Remember, however, that no amount of skillful logic design can overcome a poor microarchitecture. Once the logic has been selected, the delay can be tuned at the circuit level by choosing transistor sizes or using other styles of CMOS logic. Finally, delay is dependent on the layout. The floorplan (either manually or automatically generated) is of great importance because it determines the wire lengths that can dominate delay. Tuning of particular cells can also reduce parasitic capacitance. This section focuses on the logic and circuit optimizations of selecting the number of stages of logic, the types of gates, and the transistor sizes. Chapter 6 addresses other circuit styles and layout techniques. Chapters 10 and 11 examine design of datapath and array functional blocks. Quick delay estimation is essential to designing critical paths. Although timing analyzers or circuit simulators can compute very detailed switching waveforms and accurately predict delay, good designers cannot be dependent on simulation alone. Simulation or timing analysis only answers how fast a particular circuit operates; they do not resolve the more interesting question of how the circuit could be modified to operate faster. Many novice designers spend countless hours tweaking parameters in a circuit simulator and resimulating only to find tiny improvements. Simple models that can be applied on the back of an envelope are important to be able to rapidly estimate delay, understand its Weste04r4.fm Page 69 Monday, January 5, 2004 1:24 AM 4.2 DELAY ESTIMATION origin, and figure out how it can be reduced. This section applies the RC delay model to estimate the delay of logic gates. In general, logic gate propagation delays increase linearly with the capacitive load on the gate. We begin with a few definitions: Rise time, tr = time for a waveform to rise from 20% to 80% of its steady-state value Fall time, tf = time for a waveform to fall from 80% to 20% of its steady-state value Edge rate, trf = (tr + tf )/2 Propagation delay time, tpd = maximum time from the input crossing 50% to the output crossing 50% Contamination delay time, tcd = minimum time from the input crossing 50% to the output crossing 50% Intuitively, we know that when an input changes, the output will retain its old value for at least the contamination delay and take on its new value in at most the propagation delay. We sometimes differentiate between the delays for the output rising, tpdr /tcdr, and the output falling, tpdf /tcdf. Rise/fall times are also sometimes called slopes or edge rates. Propagation and contamination delay times are also called max-time and min-time, respectively. The gate that charges or discharges a node is called the driver and the gates and wire being driven are called the load. 4.2.1 RC Delay Models Section 2.6 developed a lumped RC model for transistors. Although transistors have complex nonlinear current-voltage characteristics, they can be approximated fairly well as a switch in series with a resistor, where the effective resistance is chosen to match the average amount of current delivered by the transistor. Transistor gate and diffusion nodes have capacitance. In this section we apply the model to estimate the delay of logic gates as the RC product of the effective driver resistance and the load capacitance. Usually, logic gates use minimum-length devices for least delay, area, and power consumption. Given this, the delay of a logic gate depends on the widths of the transistors in the gate and the capacitance of the load that must be driven. 4.2.1.1 Effective Resistance and Capacitance Recall from Section 2.6 that an nMOS transistor with width of one unit is defined to have effective resistance R. The unit-width pMOS has a higher resistance that depends on its mobility relative to the nMOS transistor. For concreteness, let us assume this resistance is 2R . Wider transistors have lower resistance. For example, a pMOS transistor of double-unit width has effective resistance R. Parallel and series transistors combine like conventional resistors. When multiple transistors are in series, their resistance is the sum of each individual resistance . When multiple transistors are in parallel, the resistance is lower if they are all ON. In many gates, the worst-case delay occurs when only one of several parallel transistors is ON. In that case, the effective resistance is just that of the single transistor. 69 Weste04r4.fm Page 70 Monday, January 5, 2004 1:24 AM 70 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Example Sketch a 3-input NAND gate with transistor widths chosen to achieve effective rise and fall resistance equal to that of a unit inverter (R). Solution: Figure 4.1 shows such a gate. The three nMOS transistors are in series so the resistance is three times that of a unit transistor. Therefore, each must be three times unit width to compensate. In other words, each transistor has resistance R/3 and the series combination has resistance R. The three pMOS transistors are in parallel. In the worst case (with one of the three inputs low), only one of the three pMOS transistors is ON. Therefore, each must be twice unit width to have resistance R. 2 2 2 3 3 3 FIG 4.1 3-input NAND gate with unit rise and fall resistance Also recall that the capacitance consists of gate capacitance and source/drain diffusion capacitance. Define the gate capacitance of a unit transistor to be Cg and the diffusion capacitance of its contacted source and drain to each be Cdiff . In many processes the capacitances are approximately equal and can be labeled C = Cg = Cdiff to keep estimation simple. The second terminal of the diffusion capacitor is the body, which is usually tied to ground (for nMOS) or VDD (for pMOS). As the DC voltage on the second terminal is irrelevant to delay, we often draw both capacitances to ground for simplicity. The gate capacitance includes fields terminating on the channel, source, and drain. To make hand analysis tractable, it can be approximated as a single capacitance to the VDD or GND rail. Capacitance scales proportionally to transistor width. 4.2.1.2 Diffusion Capacitance Layout Effects In a good layout, diffusion nodes are shared wherever possible to reduce the diffusion capacitance. Moreover, the uncontacted diffusion nodes between series transistors are usually smaller than those that must be contacted. Such uncontacted nodes have less capacitance (see Sections 2.3.3 and 5.4.4), although we will neglect the difference for hand calculations. Weste04r4.fm Page 71 Monday, January 5, 2004 1:24 AM 4.2 DELAY ESTIMATION 71 Example Annotate the 3-input NAND gate of Figure 4.1 with its gate and diffusion capacitances. Assume all diffusion nodes are contacted. Solution: Figure 4.2(a) shows the gate and its capacitances. Each input drives five units of gate capacitance. Notice that the capacitors on source diffusions attached to the rails have both terminals shorted together so they are irrelevant to circuit operation. Figure 4.2(b) redraws the gate with these capacitances deleted and the remaining capacitances lumped to ground. 2C 2 2C 2C 2C 2 2C 2 2C 3C 3C 3C (a ) 2C 2C 2 2C 3 3 3 3C 5C 5C 3C 3C 2 3 5C 3C 2 3 3 9C 3C 3C (b) FIG 4.2 3-input NAND gate annotated with capacitances In summary, gate capacitance can be determined directly from the transistor widths in the schematic. Diffusion capacitance depends on the layout. A conservative method of estimating capacitances before layout is to assume uncontacted diffusion between series transistors and contacted diffusion on all other nodes. 4.2.1.3 Elmore Delay Model Viewing ON transistors as resistors, we see that a chain of transistors can be represented as an RC ladder as shown in Figure 4.3. The Elmore delay [Elmore48] model estimates the delay of an RC ladder as the sum over each node in the ladder of the resistance between that node and a supply multiplied by the capacitance on the node: t pd = ∑ RiC i i R1 R2 R3 C1 C2 RN C3 FIG 4.3 RC ladder for Elmore delay (4.1) CN Weste04r4.fm Page 72 Monday, January 5, 2004 1:24 AM 72 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Example Figure 4.4 shows a layout of the 3-input NAND gate. A single drain diffusion region is shared between two of the pMOS transistors. Estimate the actual diffusion capacitance from the layout. Solution: Figure 4.5(a) shows the diffusion capacitance on the layout, neglecting those attached to the rails. Figure 4.5(b) redraws the schematic with these capacitances lumped to ground. Observe that two of the pMOS share a single diffusion region so the capacitance is smaller than predicted in Figure 4.2(b). FIG 4.4 3-input NAND gate layout Shared Contacted Diffusion Isolated Contacted Diffusion Merged Uncontacted Diffusion 2 2 2 3 3 3 (a) (b) FIG 4.5 3-input NAND annotated with diffusion capacitances extracted from the layout 7C 3C 3C Weste04r4.fm Page 73 Monday, January 5, 2004 1:24 AM 4.2 DELAY ESTIMATION Example Sketch a 2-input NAND gate with transistor widths chosen to achieve effective rise and fall resistance equal to a unit inverter. Compute the rising and falling propagation delays (in terms of R and C) of the NAND gate driving h identical NAND gates using the Elmore delay model. If C = 2 fF/ m and R = 2.5 k • m in a 180 nm process, what is the delay of a fanout-of-3 NAND gate? Solution: Figure 4.6(a) shows such a NAND gate annotated with diffusion capacitance assuming contacted diffusion on each transistor except between the series nMOS transistors. The two nMOS transistors are in series so each must be made twice as wide to achieve overall unit resistance. The two pMOS transistors are in parallel. In the worst case, only one is ON, so it must be the same width as in the inverter, i.e., twice unit width because pMOS transistors have lower mobility than nMOS transistors. Each input is connected to 4 units of gate width. Hence the output load of h identical NAND gates may be represented as 4h units of capacitance. Figure 4.6(b) shows the equivalent circuit for estimating the rising delay. Only one pMOS transistor is ON in the slowest case. The diffusion capacitance of the nMOS transistor is ignored because it is not on the path between the supply rail and the output node Y. This results in a slightly optimistic result. The delay is tpdr = R • ((6 + 4h)C) = (6 + 4h)RC. Figure 4.6(c) shows the equivalent circuit for estimating the falling delay. In the worst case, input A is already ‘1,’ so node x is charged up to nearly VDD through the top nMOS transistor. Input B rises, turning on the bottom nMOS transistor and thus discharging both the capacitance on node x and the output capacitance. The Elmore delay is tpdf = (R/2)(2C) + R • ((6 + 4h)C) = (7 + 4h)RC. Despite the fact that the rise and fall resistances are equal, the falling propagation delay is slightly longer than the rising delay on account of the time needed to discharge the internal parasitic capacitance. Observe that the best-case (contamination) delay of the gate could be substantially less. For example, if both inputs fell simultaneously, the output would be pulled up in half the time through the parallel combination of the two pMOS transistors: tcdr = (R/2)((6 + 4h)C) = (3 + 2h)RC. If input B were already ‘1’ and input A rises, node x would already have been discharged and thus could be ignored for delay purposes, reducing the falling delay to tcdf = R((6 + 4h)C) = (6 + 4h)RC. Hence, for fastest response, the latest input should be connected to the transistor closest to the output node when feasible. In the 180 nm process, RC = 5 ps. Therefore, a fanout-of-4 NAND gate (h = 4) has a delay of (6 + 4 • 4) • 5 ps = 110 ps. 2 2 A 2 B 2x (a ) 6C Y 4hC R Y (6+4h)C 2C (b) FIG 4.6 NAND gate delay estimation x R/2 (c) R/2 2C Y (6+4h)C 73 Weste04r4.fm Page 74 Monday, January 5, 2004 1:24 AM 74 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Example Suppose the widths of the transistors in the NAND gate are increased by a factor of k but the load is left unchanged at 4h′C (where h′= h in the previous example). In other words, the load is equivalent to h = h′/k NAND gates that are also a factor of k larger. Recompute the rising and falling propagation delays using the Elmore delay model. Solution: Figure 4.7 shows the new equivalent circuits. The Elmore delay model predicts a rising delay of (6 + 4(h′/k))RC = (6 + 4h)RC and a falling delay of (7 + 4(h′/k))RC = (7 + 4h)RC. 2k 2k A 2k B 2k (a ) 6kC 4h'C R/2k R/k 2kC (6k+4h')C (b) R/2k 2kC (6k+4h')C (c) FIG 4.7 Scaled NAND gate delay estimation Observe that the delay consists of two components. The parasitic delay of 6 or 7 is determined by the gate driving its own internal diffusion capacitance. Boosting the width of the transistors decreases the resistance but increases the capacitance so the parasitic delay is ideally independent of the gate size1. The effort delay of 4(h′/k)C = 4hC depends on the ratio (h) of external load capacitance to input capacitance and thus changes with transistor widths. The factor 4 is set by the complexity of the gate. The capacitance ratio is called the electrical effort or fanout and the term indicating gate complexity is called the logical effort. These components will be explored further in the subsequent sections. It is often helpful to express delay in a process-independent form so that circuits can be compared based on topology rather than speed of the manufacturing process. Moreover, with a process-independent measure for delay, knowledge of circuit speeds gained while working in one process can be carried over to a new process. Recall that the delay of an ideal inverter with no parasitic capacitance is τ = 3RC. We denote the normalized delay as multiples of this inverter delay: d = tpd /τ. Hence, the rising delay of the 2-input NAND gate is d = (4/3)h + 2. The RC delay model similarly predicts an inverter with real parasitics driving h identical inverters to have a delay of h + 1. 1 Of course, gates with wider transistors may use layout tricks so the diffusion capacitance increases less than linearly with width, slightly decreasing the parasitic delay of large gates. Section 8.10.1.3 illustrates folding of wide transistors. Weste04r4.fm Page 75 Monday, January 5, 2004 1:24 AM 4.2 4.2.2 DELAY ESTIMATION 75 Linear Delay Model In general the propagation delay of a gate can be written as d= f +p (4.2) where p is the parasitic delay inherent to the gate when no load is attached; f is the effort delay or stage effort that depends on the complexity and fanout of the gate: f = gh (4.3) The complexity is represented by the logical effort, g [Sutherland99]. An inverter is defined to have a logical effort of 1. More complex gates have greater logical efforts, indicating that they take longer to drive a given fanout. For example, the logical effort of the NAND gate from the previous example is 4/3. A gate driving h identical copies of itself is said to have a fanout or electrical effort of h. If the load is not identical copies of the gate, the electrical effort can be computed as C out C in where Cout is the capacitance of the external load being driven and Cin is the input capacitance of the gate2. Figure 4.8 plots normalized delay vs. electrical effort for an idealized inverter and 2-input NAND gate. The y-intercepts indicate the parasitic delay, i.e., the delay when the gate drives no load. The slope of the lines is the logical effort. The inverter has a slope of 1 by definition. The NAND has a slope of 4/3. The logical effort and parasitic delay can be estimated using RC models, as will be explored in the next sections, or extracted by curve-fitting simulated data, as discussed in Section 5.5.3. Logic gates fit the linear delay vs. fanout model remarkably well even in advanced processes; for example, Figure 5.28 shows agreement within 0.5 ps in a 180 nm process. A properly calibrated linear delay model is widely used by CAD tools such as logic synthesizers and static timing analyzers, although the notation varies from tool to tool. For example, the popular Synopsys Design Compiler tool uses the following basic model to define delay for a library of gates: (4.4) 2-input NAND 6 Normalized Delay: d h= g = 4/3 p=2 d = (4/3)h + 2 5 3 g=1 p=1 d=h+1 2 Effort Delay: f 4 1 Parastitic Delay: p 0 0 1 2 3 4 5 Electrical Effort: h = C / Cin out FIG 4.8 Normalized delay vs. fanout delay_rise = intrinsic_rise + rise_resistance * capacitance delay_fall = intrinsic_fall + fall_resistance * capacitance 2 Inverter Some TTL designers say a gate has a fanout of h when it drives h other gates even if the other gates have different capacitances. This definition would not be useful for calculating delay and is best avoided in VLSI design. The term electrical effort avoids this potential confusion and emphasizes the parallels with logical effort. Weste04r4.fm Page 76 Monday, January 5, 2004 1:24 AM 76 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Note that separate rising and falling delays are computed. These parameters are related to the logical effort terms as given in Table 4.1. The effective resistance of a gate increases with the logical effort of the gate but decreases with the gate size (i.e., input capacitance). Table 4.1 Relationship between Logical Effort and Synopsys Terminology Logical Effort Term d p Cout g/Cin Synopsys Term delay intrinsic capacitance resistance Some designers use the term drive as the reciprocal of resistance: drive = Cin/g. Large gates have greater drive. Gates with high logical effort have less drive. Delay can be expressed in terms of drive as d= 4.2.3 C out +p drive (4.5) Logical Effort Logical effort of a gate is defined as the ratio of the input capacitance of the gate to the input capacitance of an inverter that can deliver the same output current. Equivalently, logical effort indicates how much worse a gate is at producing output current as compared to an inverter, given that each input of the gate may only present as much input capacitance as the inverter. Logical effort can be measured in simulation from delay vs. fanout plots as the ratio of the slope of the delay of the gate to the slope of the delay of an inverter. Alternatively, it can be estimated by sketching gates. Figure 4.9 shows inverter, NAND, and NOR gates with transistor widths chosen to achieve unit resistance, assuming pMOS transistors have twice the resistance of nMOS transistors3. The inverter presents 3 units of input capacitance. The NAND presents 4 units of capacitance on each input, so the logical effort is 4/3. Similarly, the NOR presents 5 units of capacitance, so the logical effort 3 This assumption is made throughout the book. Exercises 4.19–4.20 explore the effects of different relative resistances (see also [Sutherland99]). The overall conclusions do not change very much, so the simple model is good enough for most hand estimates. A simulator or static timing analyzer should be used when more accurate results are required. Weste04r4.fm Page 77 Monday, January 5, 2004 1:24 AM 4.2 2 A 2 A Y 2 Y 1 (a ) A 2 B 4 B 4 2 (b) Cin = 3 g = 3/3 DELAY ESTIMATION Y 1 1 (c) Cin = 4 g = 4/3 Cin = 5 g = 5/3 FIG 4.9 Logic gates sized for unit resistance is 5/3. This matches our expectation that NANDs are better than NORs because NORs have slow pMOS transistors in series. Table 4.2 lists the logical effort of common gates. The effort tends to increase with the number of inputs. NAND gates are better than NOR gates because the series transistors are nMOS rather than pMOS. Exclusive-OR gates are particularly costly and have different logical efforts for different inputs. An interesting case is that multiplexers built from ganged tristates, as shown in Figure 1.29(b), have a logical effort of 2 independent of the number of inputs. This might at first seem to imply that very large multiplexers are just as fast as small ones. However, the parasitic delay does increase with multiplexer size; hence, it is generally fastest to construct large multiplexers out of trees of 4-input multiplexers [Sutherland99]. Table 4.2 Logical effort of common gates Gate Type 1 inverter 1 NAND NOR tristate, 2 multiplexer XOR, XNOR 4.2.4 2 Number of Inputs 3 4 4/3 5/3 2 5/3 7/3 2 6/3 9/3 2 4, 4 6, 12, 6 n (n + 2)/3 (2n + 1)/3 2 8, 16, 16, 8 Parasitic Delay The parasitic delay of a gate is the delay of the gate when it drives zero load. It can be estimated with RC delay models. A crude method good for hand calculations is to count only diffusion capacitance on the output node. For example, consider the gates in Figure 4.9, assuming each transistor on the output node has its own drain diffusion contact. Transis- 77 Weste04r4.fm Page 78 Monday, January 5, 2004 1:24 AM 78 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION tor widths were chosen to give a resistance of R in each gate. The inverter has 3 units of diffusion capacitance on the output, so the parasitic delay is 3RC = τ. In other words, the normalized parasitic delay is 1. In general, we will call the normalized parasitic delay pinv. pinv is the ratio of diffusion capacitance to gate capacitance in a particular process. It is usually close to 1 and will be considered to be 1 on many examples for simplicity. The NAND and NOR each have 6 units of diffusion capacitance on the output, so the parasitic delay is twice as great (2pinv, or simply 2). Table 4.3 estimates the parasitic delay of common gates. Increasing transistor sizes reduces resistance but increases capacitance correspondingly, so parasitic delay is, to first order, independent of gate size. However, wider transistors can be folded and often see less than linear increases in internal wiring parasitic capacitance, so in practice, larger gates tend to have slightly lower parasitic delay. Table 4.3 Parasitic delay of common gates Gate Type 1 inverter 1 NAND NOR tristate, 2 multiplexer Number of Inputs 3 4 2 2 2 4 3 3 6 4 4 8 n n n 2n This method of estimating parasitic delay is obviously crude. More refined estimates use the Elmore delay counting internal parasitics or extract the delays from simulation. The parasitic delay also depends on the ratio of diffusion capacitance to gate capacitance. For example, in a silicon-on-insulator process in which diffusion capacitance is much less, the parasitic delays will be lower. While knowing the parasitic delay is important for accurately estimating gate delay, we will see in Section 4.3 that the best transistor sizes for a particular circuit are only weakly dependent on parasitic delay. Hence, crude estimates may be acceptable. Nevertheless, it is important to realize that parasitic delay grows more than linearly with the number of inputs in a real NAND or NOR circuit. For example, Figure 4.10 shows a model of an n-input NAND gate in which the upper inputs were all ‘1’ and the bottom input rises. The gate must discharge the diffusion capacitances of all of the internal nodes as well as the output. The Elmore delay is n −1 t pd = R (3nC ) + ∑ ( i =1 n2 5 iR )(nC ) = + n RC n 2 2 (4.6) Weste04r4.fm Page 79 Monday, January 5, 2004 1:24 AM 4.2 2 2 DELAY ESTIMATION 2 n 3nC R/n n n n R/n R/n R/n nC nC nC nC nC 3nC nC FIG 4.10 n-input NAND gate parasitic delay This delay grows quadratically with the number of series transistors n, indicating that beyond a certain point it is faster to split a large gate into a cascade of two smaller gates. We will see in 4.2.5.3 that the coefficient of the n2 term tends to be even larger in real circuits than in this simple model because of gate-source capacitance. In practice, it is rarely advisable to construct a gate with more than four or possibly five series transistors. When building large fan-in gates, trees of NAND gates are better than NOR gates because the nMOS transistors have lower resistance than pMOS transistors of the same size and capacitance. 4.2.5 Limitations to the Linear Delay Model Logical effort is built on the linear delay model. Although the model works remarkably well for many practical applications, it also has limitations that should be understood when more accuracy is needed. 4.2.5.1 Input and Output Slope The largest source of error in the linear delay model is the input slope effect. Figure 4.11(a) shows a fanout-of-4 inverter driven by ramps with different slopes. Recall that the ON current increases with the gate voltage for an nMOS transistor. We say the transistor is OFF for Vgs < Vt, fully ON for Vgs = VDD, and partially ON for intermediate gate voltages. As the rise time of the input increases, the delay also increases because the active transistor is not turned all the way ON at once. Figure 4.11(b) plots average inverter propagation delay vs. input rise time. Notice that the delay vs. rise time data fits a straight line quite well. [Hedenstierna87] suggests that the line may be modeled as t pd = t pd − step Vt 1 + 2V DD + t edge 6 (4.7) where tpd–step is the propagation delay assuming a step input and tedge is the appropriate edge rate (tr or tf ). 79 Weste04r4.fm Page 80 Monday, January 5, 2004 1:24 AM 80 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Vin Vout 4x tr Vout 100 t r = 150 ps 1.5 80 t r = 50 ps t r = 250 ps 60 1.0 t pd 40 0.5 20 Vin 0.0 0.0 (a) 0 t (ps) 100 200 300 400 500 0 (b) 50 100 150 200 250 tr FIG 4.11 SPICE simulation of slope effect on CMOS inverter delay A first-order RC circuit has a 20%-80% rise/fall time of (ln 0.8 – ln 0.2)RC = 1.39RC. The 20%–80% rise/fall time of a gate is roughly 1-1.5x the propagation delay. Hence, the input slope of a gate is related to the propagation delay of the previous gate. The Synopsys delay model supports considering the effect of slope on gate delay by introducing a slope_rise / slope_fall term. The overall delay of a gate becomes linearly dependent on the delay of the previous gate. delay_rise = intrinsic_rise + rise_resistance • capacitance + slope_rise • delay_previous delay_fall = intrinsic_fall + fall_resistance • capacitance + slope_fall • delay_previous Accounting for slopes is important for accurate timing analysis, but is generally more complex than is worthwhile for hand calculations. Fortunately, we will see in Section 4.3 that circuits are fastest when each gate has the same effort delay and when that delay is roughly 4τ. Because slopes are related to edge rate, fast circuits tend to have relatively consistent slopes. If a cell library is characterized with these slopes, it will tend to be used in the regime in which it most accurately models delay. 4.2.5.2 Input Arrival Times Another source of error in the linear delay model is the assumption that one input of a multiple-input gate switches while the others are completely stable. When two inputs to a series stack turn ON simultaneously, the delay will be slightly longer than predicted because both transistors are only partially ON during the initial part of the transition. When two inputs to a parallel stack turn ON simulta- Weste04r4.fm Page 81 Monday, January 5, 2004 1:24 AM 4.2 DELAY ESTIMATION neously, the delay will be shorter than predicted because both transistors deliver current to the output. Figure 4.12 shows plots the propagation delay of an FO3 2-input NAND gate as a function of the input interarrival time. Input A switches at time 0, while input B switches in the same direction at time tb. Propagation delay is measured from the latest input rising for falling outputs and from the earliest input falling for rising outputs. When one input arrives well before the other, |tb| is large and the propagation delay is essentially independent of tb. This is the case assumed in the linear delay model. The delays are slightly different depending on which input arrives first, as will be explored in Section 5.5.3. When the two inputs arrive at nearly the same time, |tb| is small. tpdf increases because of the series pulldowns while tpdr decreases because of the parallel pullups. t pdf 100 80 60 2 2 40 Y B 2 A 2 20 0 –200 –150 –100 –50 0 50 100 150 200 t b (ps) FIG 4.12 Delay sensitivity to input arrival time 4.2.5.3 Gate-Source Capacitance The examples in Section 4.2.1 assumed that gate capacitance terminates on a fixed supply rail. As discussed in Section 2.3.2, the bottom terminal of the gate oxide capacitor is the channel, which is primarily connected to the source when the transistor is ON. This means that as the source of a transistor changes value, charge is required to change the voltage on Cgs. Figure 4.13(a) revisits the 2-input NAND gate example, explicitly showing gate-source capacitances. As node x is discharged on a falling output transition in Figure 4.13(b), Cgs2 must also be discharged. The delay can now be estimated as (R/2)(2C + 2C) + R((2 + 2 + 2 + 4h)C) = (8 + 4h)RC. RC models are more valuable for simplicity than accuracy, so some designers ignore this effect in hand calculations. Note that Cgs of the pMOS transistors does not affect delay because it is not on the path between Y and GND. However, the gate capacitance of both nMOS and pMOS would affect the loading and delay of the previous stage. 81 Weste04r4.fm Page 82 Monday, January 5, 2004 1:24 AM 82 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION 2 2 A 2C B 2C 6C 2 2x Y 4hC R/2 x R/2 2C 2C Y (6+4h)C 2C (a ) (b) FIG 4.13 NAND gate delay estimation with gate-source capacitance modeled 4.2.5.4 Bootstrapping Transistors also have some capacitance from gate to drain. This capacitance couples the input and output in an effect known as bootstrapping, which can be understood by examining Figure 4.14(a). Our models so far have only considered Cin (Cgs). This figure also considers Cgd, the gate to drain capacitance. In the case that the input is rising (the output starts high), the effective input capacitance is Cgs + Cgd. When the output starts to fall, the voltage across Cgd changes, requiring the input to supply additional current to charge Cgd. In other words, the impact of Cgd on gate capacitance is effectively doubled. To illustrate the effect of the bootstrap capacitance on a circuit, Figure 4.14(b) shows two inverter pairs. The top pair has an extra bit of capacitance between the input and output of the second inverter. The bottom pair has the same amount of extra capacitance from input to ground. When x falls, nodes a and c begin to rise (Figure 4.14(c)). At first, both nodes see approximately the same capacitance, consisting of the two transistors and the extra 3 fF. As node a rises, it initially bumps up b or “lifts b by its own bootstraps.” Eventually the nMOS transistors turn ON, pulling down b and d. As b falls, it tugs on a 32/2 a 32/2 b c a 3 fF 16/2 16/2 x Vin Cgd x 32/2 c Vout 16/2 Cgs (a ) 32/2 d d 16/2 3 fF (b) b (c) FIG 4.14 The effect of bootstrapping on inverter delay and waveform shape Weste04r4.fm Page 83 Monday, January 5, 2004 1:24 AM 4.3 LOGICAL EFFORT AND TRANSISTOR SIZING through the capacitor, leading to the slow final transition visible on node a. Also observe that b falls later than d because of the extra charge that must be supplied to discharge the bootstrap capacitor. In summary, the extra capacitance has a greater effect when connected between input and output as compared to when it is connected between input and ground. Because Cgd is fairly small, bootstrapping is only a mild annoyance in digital circuits. However, if the inverter is biased in its linear region near VDD /2, the Cgd may appear multiplied by the gain of the inverter. This is known as the Miller effect and is of major importance in analog circuits. 4.3 Logical Effort and Transistor Sizing Designers often need to choose the fastest circuit topology and gate sizes for a particular logic function and to estimate the delay of the design. As has been stated, simulation or timing analysis are poor tools for this task because they only determine how fast a particular implementation will operate, not whether the implementation can be modified for better results and if so, what to change. Inexperienced designers often end up in the “simulate and tweak” loop involving minor changes and many simulations. This is not only tedious but also seldom results in significant improvements. The method of Logical Effort [Sutherland99] provides a simple method “on the back of an envelope” to choose the best topology and number of stages of logic for a function. It allows the designer to quickly estimate the minimum possible delay for the given topology and to choose gate sizes that achieve this delay. Logical Effort is based on the linear delay model. We first review using the model to estimate the delay of individual logic gates. The method generalizes to predict the delay of multistage logic networks and to choose the best number of stages for a multistage network. This section concludes with an example applying Logical Effort to design a memory decoder and summarizes the key insights from the method. The techniques of Logical Effort will be revisited throughout this text to understand delay of many types of circuits. 4.3.1 Delay in a Logic Gate The linear delay model of EQs (4.2), (4.3), and (4.4) express propagation delay of a logic gate in terms of the complexity of a gate (its logical effort, g), the capacitive fanout (electrical effort, h), and the parasitic delay, p. Let us begin with two examples. 83 Weste04r4.fm Page 84 Monday, January 5, 2004 1:24 AM 84 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Example Estimate the delay of the fanout-of-4 (FO4) inverter (i.e., an inverter driving four identical copies) shown in Figure 4.15. Assume the inverter is constructed in a 180 nm process with τ = 15 ps. Solution: The logical effort of the inverter is g = 1, by definition. The electrical effort is 4 because the load is four gates of equal size. The parasitic delay of an inverter is pinv ≈ 1. The total delay is d = gh + p = 1 • 4 + 1 = 5 in normalized terms, or tpd = 75 ps in absolute terms. Often path delays are expressed in terms of FO4 inverter delays. While not all designers are familiar with the τ notation, most experienced designers do know the delay of a fanout-of-4 inverter in the process in which they are working. τ can be estimated as 0.2 FO4 inverter delays. Even if the ratio of diffusion capacitance to gate capacitance changes so pinv = 0.8 or 1.2 rather than 1, the FO4 inverter delay only varies from 4.8 to 5.2. Hence, the delay of a gate-dominated logic block expressed in terms of FO4 inverters remains relatively constant from one process to another even if the diffusion capacitance does not. As a rough rule of thumb, the FO4 delay for a d process (in picoseconds) is 1/3 to 1/2 of the channel length (in nanometers). For example, a 180 nm process may have an FO4 delay of 60-90 ps. Delay is highly sensitive to process, voltage, and temperature variations, as will be examined in Section 6.6. The FIG 4.15 Fanout-of-4 FO4 delay is usually quoted assuming typical process (FO4) inverter parameters and worst-case environment (low power supply voltage and high temperature). 4.3.2 Delay in Multistage Logic Networks Logical Effort generalizes to multistage logic networks. For example, Figure 4.16 shows the logical and electrical efforts of each stage in a multistage path as a function of the sizes of each stage. The path of interest (the only path in this case) is marked with the dashed blue line. Observe that logical effort is independent of size, while electrical effort depends on sizes. 10 g1 = 1 h1 = x/10 x g2 = 5 /3 h2 = y/x y g3 = 4 /3 h3 = z /y FIG 4.16 Multistage logic network z g4 = 1 h4 = 2 0/z 20 Weste04r4.fm Page 85 Monday, January 5, 2004 1:24 AM 4.3 LOGICAL EFFORT AND TRANSISTOR SIZING Example A ring oscillator is constructed from an odd number of inverters, as shown in Figure 4.17. Estimate the frequency of an N-stage ring oscillator. Solution: The logical effort of the inverter is g = 1, by definition. The electrical effort of each inverter is also 1 because it drives a single identical load. The parasitic delay is also 1. The delay of each stage is d = gh + p = 1 • 1 + 1 = 2. An N-stage ring oscillator has a period of 2N stage delays because a value must propagate twice around the ring to regain original polarity. Therefore, the period is T = 2 • 2N. The frequency is the reciprocal of the period, 1/4N. A 31-stage ring oscillator in a 180 nm process has a frequency of 1/(4 • 31 • 15 ps) = 540 MHz. Note that ring oscillators are often used as process monitors to judge if a particular chip is faster or slower than nominally expected. One of the inverters should be replaced with a NAND gate to turn the ring off when not in use. The output can be routed to an external pad, possibly through a test multiplexer. The oscillation frequency should be low enough (e.g., 100 MHz) that the path to the outside world is not a limiter. FIG 4.17 Ring oscillator The path logical effort G can be expressed as the products of the logical efforts of each stage along the path. G = ∏ gi (4.8) The path electrical effort H can be given as the ratio of the output capacitance the path must drive divided by the input capacitance presented by the path. This is more convenient than defining path electrical effort as the product of stage electrical efforts because we do not know the individual stage electrical efforts until gate sizes are selected. H= C out(path) C in(path) (4.9) 85 Weste04r4.fm Page 86 Monday, January 5, 2004 1:24 AM 86 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION The path effort F is the product of the stage efforts of each stage. Can we by analogy state F = GH? F = ∏ f i = ∏ gi hi 15 90 5 15 90 FIG 4.18 Circuit with two-way branch (4.10) In paths that branch, F ≠ GH. This is illustrated in Figure 4.18, a circuit with a two-way branch. Consider a path from the primary input to one of the outputs. The path logical effort is G = 1 • 1 = 1. The path electrical effort is H = 90/5 = 18. Thus, GH = 18. But F = f 1 f 2 = g 1 h 1 g 2 h 2 = 1 • 6 • 1 • 6 = 36. In other words, F = 2GH in this path on account of the two-way branch. We must introduce a new kind of effort to account for branching between stages of a path. This branching effort b is the ratio of the total capacitance seen by a stage to the capacitance on the path; in Figure 4.18 it is (15+15)/15 = 2. b= C onpath + C offpath C onpath (4.11) The path branching effort B is the product of the branching efforts between stages. B = ∏ bi (4.12) Now we can define the path effort F as the product of the logical, electrical, and branching efforts of the path. Note that the product of the electrical efforts of the stages is actually BH, not just H. F = GBH (4.13) We can now compute the delay of a multistage network. The path delay D is the sum of the delays of each stage. It can also be written as the sum of the path effort delay DF and path parasitic delay P. D = ∑ di = DF + P DF = ∑ f i (4.14) P = ∑ pi The product of the stage efforts is F, independent of gate sizes. The path effort delay is the sum of the stage efforts. The sum of a set of numbers whose product is constant is Weste04r4.fm Page 87 Monday, January 5, 2004 1:24 AM 4.3 LOGICAL EFFORT AND TRANSISTOR SIZING Example Estimate the minimum delay of the path from A to B in Figure 4.19 and choose transistor sizes to achieve this delay. The initial NAND2 gate may present a load of 8 λ of transistor width on the input and the output load is equivalent to 45 λ of transistor width. Solution: The path logical effort is G = (4/3) • (5/3) • (5/3) = 100/27. The path electrical effort is H = 45/8. The path branching effort is B = 3 • 2 = 6. The path effort is F = GBH = 125. As there are three stages, the best stage effort is ˆ f = 3 125 = 5 . The path parasitic delay is P = 2 + 3 + 2 = 7. Hence, the minimum path delay is D = 3 • 5 + 7 = 22 in units of τ, or 5.4 FO4 inverter delays. The gate sizes are computed with the capacitance transformation: y = 45 • (5/3)/5 = 15. x = (15 + 15) • (5/3)/5 = 10. We check that the initial 2-input NAND gate should have a size of (10 + 10 + 10) • (4/3)/5 = 8, as desired. The transistor sizes in Figure 4.20 are chosen to give the desired amount of input capacitance while achieving equal rise and fall delays. We can also check that our delay was achieved. The NAND2 gate delay is d1 = g1h1 + p1 = (4/3) • (10 + 10 + 10)/8 + 2 = 7. The NAND3 gate delay is d2 = g2h2 + p2 = (5/3) • (15 + 15)/10 + 3 = 8. The NOR2 gate delay is d3 = g3h3 + p3 = (5/3) • 45/15 + 2 = 7. Hence, the path delay is 22, as predicted. x Many inexperienced designers y know that wider transistors offer x 45 more current and thus try to make circuits faster by using bigger gates. A 8 x y B Increasing the size of any of the gates 45 except the first one only makes the circuit slower. For example, increasing the size of the NAND3 makes the FIG 4.19 Example path NA N D 3 f a s t e r b u t m a k e s t h e NAND2 slower, resulting in a net speed loss. Increasing the size of the initial NAND2 gate does speed up the circuit under consideration. How45 ever, it presents a larger load on the path that computes input A, making A P: 4 P: 4 P: 12 N: 4 B that path slower. Hence, it is crucial N: 6 45 N: 3 to have a specification of not only the load the path must drive but also the maximum input capacitance the path FIG 4.20 Example path annotated with transistor sizes may present. 87 Weste04r4.fm Page 88 Monday, January 5, 2004 1:24 AM 88 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION minimized by choosing all the numbers to be equal. In other words, the path delay is minimized when each stage bears the same effort fˆ . If a path has N stages and each bears the same effort, that effort must be fˆ = gi hi = F 1 / N (4.15) Thus the minimum possible delay of an N-stage path with path effort F and path parasitic delay P is D = NF 1/ N + P (4.16) This is a key result of Logical Effort. It shows that the minimum delay of the path can be estimated knowing only the number of stages, path effort, and parasitic delays without the need to assign transistor sizes. This is much superior to simulation, in which delay depends on sizes and you never achieve certainty that the sizes selected are those that offer minimum delay. It is also straightforward to select gate sizes to achieve this least delay. Combining EQs (4.3) and (4.4) gives us the capacitance transformation formula to find the best input capacitance for a gate given the output capacitance it drives. C in i = C out i • gi ˆ f (4.17) Starting with the load at the end of the path, work backward applying the capacitance transformation to determine the size of each stage. Check the arithmetic by verifying that the size of the initial stage matches the specification. 4.3.3 Choosing the Best Number of Stages Given a specific circuit topology, we now know how to estimate delay and choose gate sizes. However, there are many different topologies that implement a particular logic function. Logical Effort tells us that NANDs are better than NORs and that gates with few inputs are better than gates with many. In this section we will also use Logical Effort to predict the best number of stages to use. Logic designers sometimes estimate delay by counting the number of stages of logic, assuming each stage has a constant “gate delay.” This is potentially misleading because it implies that the fastest circuits are those that use the fewest stages of logic. Of course the gate delay actually depends on the electrical effort, so sometimes using fewer stages results in more delay. The following example illustrates this point. In general, you can always add inverters to the end of a path without changing its function (save possibly for polarity). Let us compute how many should be added for least delay. The logic block shown in Figure 4.22 has n1 stages and a path effort of F. Consider adding N – n1 inverters to the end to bring the path to N stages. The extra inverters do not Weste04r4.fm Page 89 Monday, January 5, 2004 1:24 AM 4.3 LOGICAL EFFORT AND TRANSISTOR SIZING 89 Example A control unit generates a signal from a unit-sized inverter. The signal must drive unit-sized loads in each bitslice of a 64-bit datapath. The designer can add inverters to buffer the signal to drive the large load. Assuming polarity of the signal does not matter, what is the best number of inverters to add and what delay can be achieved? Solution: Figure 4.21 shows the cases of adding 0, 1, 2, or 3 inverters. The path electrical effort is H = 64. The path logical effort is G = 1, independent of the number of inverters. Thus the path effort is F = 64. The inverter sizes are chosen to achieve equal stage effort. The total delay is D = N N 64 + N . The 3-stage design is fastest and is much superior to a single stage. If an even number of inversions were required, the two- or four-stage designs are promising. The four-stage design is slightly faster, but the two-stage design requires significantly less area and power. 1 1 1 1 8 4 2.8 16 Initial Driver 8 23 Datapath Load 64 N: f: D: 1 64 65 64 2 8 18 64 64 3 4 15 4 2.8 15.3 Fastest FIG 4.21 Comparison of different number of stages of buffers change the path logical effort but do add parasitic delay. The delay of the new path is n1 D = NF 1 / N + ∑ pi + ( N − n1 ) p inv i =1 Logic Block n1 Stages Path Effort F N – n1 Extra Inverters (4.18) FIG 4.22 Logic block with additional inverters Weste04r4.fm Page 90 Monday, January 5, 2004 1:24 AM CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Differentiating with respect to N and setting to 0 allows us to solve for the best number of stages. The result can be expressed more compactly by defining ˆ ρ = F 1/ N to be the best stage effort. ∂D = − F 1 / N ln F 1 / N + F 1 / N + p inv = 0 ∂N ⇒ p inv + ρ(1 − ln ρ) = 0 (4.19) EQ (4.19) has no closed form solution. Neglecting parasitics (i.e., assuming pinv = 0), we find the classic result that ρ = 2.71828 (e) [Mead80]. In practice, the parasitic delays mean each inverter is somewhat more costly to add. As a result, it is better to use fewer stages, or equivalently a higher stage effort than e. Solving numerically, when pinv = 1, we find ρ = 3.59. ˆ A path achieves least delay by using N = log ρ F stages. It is important to understand not only the best stage effort and number of stages but also the sensitivity to using a different number of stages. Figure 4.23 plots the delay increase using a particular number of stages against the number of stages, for pinv = 1. The curve is very flat around the optimum. The delay is within 15% of the best achievable if the number of stages is within 2/3 to 1.5 times the theoretical best number (i.e., ρ is in the range of 2.4 to 6). 1.6 D(N) / D(N) 90 1.51 1.4 1.26 1.2 1.15 1.0 (r = 2.4) (r = 6) 0.0 0.5 0.7 1.0 N/N FIG 4.23 Sensitivity of delay to number of stages 1.4 2.0 Weste04r4.fm Page 91 Monday, January 5, 2004 1:24 AM 4.3 LOGICAL EFFORT AND TRANSISTOR SIZING Using a stage effort of 4 is a convenient choice and simplifies mentally choosing the best number of stages. This effort gives delays within 2% of minimum for pinv in the range of 0.7 to 2.5. This further explains why a fanout-of-4 inverter has a “representative” logic gate delay. 4.3.4 Example Consider a larger example to illustrate the application of Logical Effort. Our esteemed colleague Ben Bitdiddle is designing a decoder for a register file in the Motoroil 68W86, an embedded processor for automotive applications. The decoder has the following specifications: 16-word register file 32-bit words Each register bit presents a load of 3 unit-sized transistors on the word line (2 unit-sized access transistors plus some wire capacitance) True and complementary versions of the address bits A[3:0] are available Each address input can drive 10 unit-sized transistors As we will see further in Section 11.2.2, a 2N-word decoder consists of 2N N-input AND gates. Therefore, the problem is reduced to designing a suitable 4-input AND gate. Let us help Ben determine how many stages to use, how large each gate should be, and how fast the decoder can operate. The output load on a word line is 32 bits with 3 units of capacitance each, or 96 units. Therefore, the path electrical effort is H = 96/10 = 9.6. Each address is used to compute half of the 16 word lines; its complement is used for the other half. Therefore, a B = 8-way branch is required somewhere in the path. Now we are faced with a chicken-and-egg dilemma. We need to know the path logical effort to calculate the path effort and best number of stages. However, without knowing the best number of stages, we cannot sketch a path and determine the logical effort for that path. There are two ways to resolve the dilemma. One is to sketch a path with a random number of stages, determine the path logical effort, then use that to compute the path effort and the actual number of stages. The path can be redesigned with this number of stages, refining the path logical effort. If the logical effort changes significantly, the process can be repeated. Alternatively, we know that the logic of a decoder is rather simple so we can ignore the logical effort (assume G = 1). Then we can proceed with our design, remembering that the best number of stages is likely slightly higher than predicted because we neglected logical effort. Taking the second approach, we find the path effort is F = GBH = (1)(8)(9.6) = 76.8. Targeting a best stage effort of ρ = 4, we find the best number of stages is N = log 4 76.8 = 3.1 . Let us select a 3-stage design, recalling that a 4-stage design might be a good choice too when logical effort is considered. Figure 4.24 shows a possible 3-stage design (INV-NAND4-INV). 91 Weste04r4.fm Page 92 Monday, January 5, 2004 1:24 AM 92 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION A[3] A[3] 10 A[2] A[2] 10 10 10 A[1] A[1] 10 10 A[0] A[0] 10 10 y z word[0] 96 units of wordline capacitance y z word[15] FIG 4.24 3-stage decoder design The path has a logical effort of G = 1 • (6/3) • 1 = 2, so the actual path effort is F = ˆ (2)(8)(9.6) = 154. The stage effort is f = 154 1 / 3 = 5.36 . This is in the reasonable range of 2.4 to 6, so we expect our design to be acceptable. Applying the capacitance transformation, we find gate sizes z = 96 • 1/5.36 = 18 and y = 18 • 2 /5.36 = 6.7. The delay is 3 • 5.36 + 1 + 4 + 1 = 22.1. Logical Effort also allows us to rapidly compare alternative designs using a spreadsheet rather than a schematic editor and a large number of simulations. Table 4.4 compares a number of alternative designs. We find a 4-stage design is somewhat faster, as we suspected. The 4-stage NAND2-INV-NAND2-INV design not only has the theoretical best number of stages but also uses simpler 2-input gates to reduce the logical effort and parasitic delay to obtain a 12% speedup over the original design. However, the 3-stage design has a smaller total gate area and dissipates less power. Table 4.4 Spreadsheet comparing decoder designs Design NAND4-INV NAND2-NOR2 INV-NAND4-INV NAND4-INV-INV-INV NAND2-NOR2-INV-INV NAND2-INV-NAND2-INV INV-NAND2-INV-NAND2-INV NAND2-INV-NAND2-INV-INV-INV Stages N G P D 2 2 3 4 4 4 5 6 2 20/9 2 2 20/9 16/9 16/9 16/9 5 4 6 7 6 6 7 8 29.8 30.1 22.1 21.1 20.5 19.7 20.4 21.6 Weste04r4.fm Page 93 Monday, January 5, 2004 1:24 AM 4.3 4.3.5 LOGICAL EFFORT AND TRANSISTOR SIZING Summary and Observations Logical Effort provides an easy way to compare and select circuit topologies, choose the best number of stages for a path, and estimate path delay. The notation takes some time to become natural but this author has poured through all the letters in the English and Greek alphabets without finding better notation. It may help to remember d for “delay,” p for “parasitic,” b for “ branching,” f for “effort,” g for “logical effort,” (or perhaps gain), and h as the next letter after “f ” and “g.” The notation is summarized in Table 4.5 for both stages and paths. Table 4.5 Summary of Logical Effort notation Term Stage Expression Path Expression number of stages 1 N logical effort g (see Table 4.2) G = ∏ gi electrical effort h= branching effort b= effort f = gh F = GBH effort delay f DF = ∑ f i parasitic delay p (see Table 4.3) P = ∑ pi delay d=f+p D = ∑ di = DF + P C out C in H= C onpath + C offpath C onpath Cout(path) Cin (path) B = ∏ bi The method of Logical Effort is applied with the following steps: 1. Compute the path effort: 2. Estimate the best number of stages: F = GBH ˆ N = log 4 F 93 Weste04r4.fm Page 94 Monday, January 5, 2004 1:24 AM 94 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION 3. Sketch a path using: ˆ N stages. 4. Estimate the minimum delay: ˆˆ D = NF 1 / N + P 5. Determine the best stage effort: fˆ = F 1 / N 6. Starting at the end, work backward to find sizes: C in i = C out i • gi fˆ CAD tools are very fast and accurate at evaluating complex delay models, so Logical Effort should not be used as a replacement for such tools. Rather, its value arises from “quick and dirty” hand calculations and from the insights it lends to circuit design. Some of the key insights include: The idea of a numeric “logical effort” that characterizes the complexity of a logic gate or path allows you to compare alternative circuit topologies and show that some topologies are better than others. NAND structures are faster than NOR structures in complementary CMOS circuits. Paths are fastest when the effort delays of each stage are about the same and when these delays are close to four. Path delay is insensitive to modest deviations from the optimum. Stage efforts of 2.4–6 give designs within 15% of minimum delay. There is no need to make calculations to more than 1–2 significant figures, so many estimations can be made in your head. There is no need to choose transistor sizes exactly according to theory and there is little benefit in tweaking transistor sizes if the design is reasonable. Using stage efforts somewhat greater than 4 reduces area and power consumption at a slight cost in speed. Using efforts greater than 6-8 comes at a significant cost in speed. Using fewer stages for “less gate delays” does not make a circuit faster. Making gates larger also does not make a circuit faster; it only increases the area and power consumption. The delay of a well-designed path is about log4F fanout-of-4 (FO4) inverter delays. Each quadrupling of the load adds about one FO4 inverter delay to the path. Control signals fanning out to a 64-bit datapath therefore incur an amplification delay of about 3 FO4 inverters. The logical effort of each input of a gate increases through no fault of its own as the number of inputs grows. Considering both logical effort and parasitic delay, we find a practical limit of about 4 series transistors in logic gates and about 4 inputs to multiplexers. Beyond this fan-in, it is faster to split gates into multiple stages of skinnier gates. Weste04r4.fm Page 95 Monday, January 5, 2004 1:24 AM 4.3 LOGICAL EFFORT AND TRANSISTOR SIZING Inverters or 2-input NAND gates with low logical efforts are best for driving nodes with a large branching effort. Use small gates after the branches to minimize load on the driving gate. W hen a path forks and one leg is more critical than the others, buffer the noncritical legs to reduce the branching effort on the critical path 4.3.6 Limitations of Logical Effort Logical Effort is based on the linear delay model and the simple premise that making the effort delays of each stage equal minimizes path delay. This simplicity is the method’s greatest strength but also results in a number of limitations: The linear delay model fails to capture the effect of input slope. Fortunately, edge rates tend to be about equal in well-designed circuits with equal effort delay per stage. The RC delay model neglects the effects of velocity saturation and overestimates the logical effort of NAND structures. It also ignores the body effect. Logical effort may be more accurately characterized through simulation, as shown in Section 5.5.3. Logical Effort does not account for interconnect. The effects of nonnegligible wire capacitance and RC delay will be revisited in Section 4.5. Logical Effort is most applicable to high-speed circuits with regular layouts where routing delay does not dominate. Such structures include adders, multipliers, memories, and other datapaths and arrays. Logical Effort explains how to design a path for maximum speed but not how to design for minimum area or power given a fixed speed constraint. Paths with complex branching are difficult to analyze by hand. 4.3.7 Extracting Logical Effort from Datasheets When using a standard cell library, you can often extract logical effort of gates directly from the datasheets. For example, Figure 4.25 shows the INV and NAND2 datasheets from the Artisan Components library for the TSMC 180 nm process. The gates in the library come in various drive strengths. INVX1 is the unit inverter; INVX2 has twice the drive. INVXL has the same area as the unit inverter but uses smaller transistors to reduce power consumption on noncritical paths. The X12-X20 inverters are built from three stages of smaller inverters to give high drive strength and low input capacitance at the expense of greater parasitic delay. From the datasheet, we see the unit inverter has an input capacitance of 3.6fF. The rising and falling delays are specified separately. We will develop a notation for different 95 Weste04r4.fm Page 96 Monday, January 5, 2004 1:24 AM 96 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION delays in Section 6.2.1.5, but will use the average delay for now. The average intrinsic or parasitic delay is (25.3 + 14.6)/2 = 20.0 ps. The slope of the delay vs. load capacitance curve is the average of the rising and falling Kload values. An inverter driving a fanout of h will thus have a delay of ( ) 4.53 + 2.37 fF tpd = 20.0 ps + 3.6 gate (h gates) 2 ns pF = ( 20.0 + 12.4h ) ps (4.20) The slope of the delay vs. fanout curve indicates τ = 12.4 ps and the y-intercept indicates pinv = 20.0 ps, or 1.61 in normalized terms. By a similar calculation, we find the X1 2-input NAND gate has an average delay from the inner (A) input of ( ) 31.3 + 19.5 4.53 + 2.84 fF tpd = ps + 4.2 gate (h gates) 2 2 ns pF = ( 25.4 + 15.5h ) ps (4.21) Thus, the parasitic delay is 2.05 and the logical effort is 1.25. The parasitic delay from the outer (B) input is slightly higher, as expected. The parasitic delay and logical effort of the X2 and X4 gates are similar, confirming our model that logical effort should be independent of gate size for gates of reasonable sizes. 4.4 Power Dissipation Static CMOS gates are very power-efficient because they dissipate nearly zero power while idle. For much of the history of CMOS design, power was a secondary consideration behind speed and area for many chips. As transistor counts and clock frequencies have increased, power consumption has skyrocketed and now is a primary design constraint. We begin by reviewing some definitions. The instantaneous power P(t) drawn from the power supply is proportional to the supply current iDD(t) and the supply voltage VDD P (t ) = i DD (t )V DD (4.22) The energy consumed over some time interval T is the integral of the instantaneous power T E = ∫ i DD (t )V DD dt 0 (4.23) Weste04r4.fm Page 97 Monday, January 5, 2004 1:24 AM 4.4 POWER DISSIPATION 97 FIG 4.25 Artisan Components cell library datasheets Weste04r4.fm Page 98 Monday, January 5, 2004 1:24 AM 98 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION The average power over this interval is T Pavg = E1 = i DD (t )V DD dt T T∫ 0 (4.24) Power dissipation in CMOS circuits comes from two components: Static dissipation due to ○ subthreshold conduction through OFF transistors ○ tunneling current through gate oxide ○ leakage through reverse-biased diodes ○ contention current in ratioed circuits Dynamic dissipation due to ○ charging and discharging of load capacitances ○ “short-circuit” current while both pMOS and nMOS networks are partially ON Ptotal = Pstatic + Pdynamic (4.25) This section quantifies each of these components of power and discusses techniques to minimize power consumption. 4.4.1 Static Dissipation Considering the static CMOS inverter shown in Figure 4.26, if the input = ‘0,’ the associated nMOS transistor is OFF and the pMOS transistor is ON. The output voltage is VDD or logic ‘1.’ When the input = ‘1,’ the associated nMOS transistor is ON and the pMOS transistor is OFF. The output voltage is 0 volts (GND). Note that one of the transistors is always OFF when the gate is in either of these logic states. Ideally, no current flows through the OFF transistor so the power dissipation is pMOS pMOS zero when the circuit is quiescent, i.e., when no transistors are switching. Zero 0 1 1 0 quiescent power dissipation is a principle advantage of CMOS over competing nMOS nMOS transistor technologies. However, secondary effects including subthreshold conduction, tunneling, and leakage lead to small amounts of static current flowing through the OFF transistor. Assuming the leakage current is constant so instanFIG 4.26 CMOS inverter model for taneous and average power are the same, the static power dissipation is the static power dissipation evaluation product of total leakage current and the supply voltage. Pstatic = I staticVDD (4.26) Weste04r4.fm Page 99 Monday, January 5, 2004 1:24 AM 4.4 POWER DISSIPATION Example A digital system in a 1.2 V 100 nm process [Parihar01] has 200 million transistors, of which 20 million are in logic gates and the remainder in memory arrays. The average logic transistor width is 12 λ and the average memory transistor width is 4 λ. The process has two threshold voltages and two oxide thicknesses. Subthreshold leakage for OFF devices is 20 nA/µm for low-threshold devices and 0.02 nA/ µm for high-threshold devices. Gate leakage is 3 nA/µm for thin oxides and 0.002 nA/µm for thick oxides. Memories use low-leakage devices everywhere. Logic uses low-leakage devices in all but 20% of the paths that are most critical for performance. Diode leakage is negligible. Estimate the static power consumption. How would the power consumption change if the low-leakage devices were not available? Solution: There are (20 • 106 logic transistors) • (0.2) • (12 λ • (0.05 µm/λ) = 2.4 • 106 µm of high-leakage devices and [(20 • 106 logic transistors) • (0.8) • (12 λ) + (180 • 106 memory transistors) • (4 λ)] • (0.05 µm/λ) = 45.6 • 106 µm of low-leakage devices. All devices exhibit gate leakage. On average, half the transistors are OFF and contribute subthreshold leakage. Therefore, the total static current is (2.4 • 106 µm) • [(20 nA/µm)/2 + (3 nA/µm)] + (45.6 • 106 µm) • [(0.02 nA/µm)/2 + (0.002 nA/µm)] = 32 mA. Static power consumption is (32 mA) • (1.2 V ) = 38 mW. This is likely to be small compared to dynamic power consumption, yet large enough to limit the battery life of battery-powered systems on standby. If low-leakage devices were not available, the total static current would be (2.4 • 106 µm + 45.6 • 106) • [(20 nA/µm)/2 + (3 nA/µm)] = 624 mA, for standby power of (624 mA) • (1.2 V) = 749 mW. According to EQ (2.34), OFF transistors still conduct a small amount of subthreshold current. As subthreshold current is exponentially dependent on threshold voltage, it is increasing dramatically as threshold voltages have scaled down. SiO2 is a very good insulator, so leakage current through the gate dielectric historically was very low. However, it is possible for electrons to tunnel across very thin insulators; the probability drops off exponentially with oxide thickness. Tunneling current becomes important for transistors around the 130 nm generation with gate oxides of 20Å or thinner. There is also some small static dissipation due to reverse biased diode leakage between diffusion regions, wells, and the substrate, as shown for an inverter in Figure 2.19. Diode leakage is given by EQ (2.38). In modern processes, diode leakage is generally much smaller than the subthreshold or gate leakage and may be neglected. In older processes, all three components of static power dissipation were small enough that CMOS was often said to consume “zero” DC power. Leakage power was of concern only to ultra-low-power systems. In 130 nm processes and beyond, the static power is rapidly becoming a primary design issue and vendors now provide leakage data, often in the form of nA/µm of gate length. Hand-held battery-powered devices typically require 99 Weste04r4.fm Page 100 Monday, January 5, 2004 1:24 AM 100 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Example µA/V2 , For a process with k ′p of 75 Vtp= –0.4 V, and VDD = 1.8 V, calculate the static power dissipation of a 32 word x 48-bit ROM that contains a 1:32 pseudo-nMOS row decoder and pMOS pull-ups on the 48-bit lines. The W/L ratio of the pMOS pull-ups is 1. Assume one of the word lines and 50% of the bitlines are high at any given time. Solution: Each pMOS transistor dissipates power when the output is low. I pull-up Ppull-up ( ) 2 W V DD − V tp = = 73µA L 2 = V DD I pull-up = 130µW k'p (4.27) We expect to see 31 wordlines and 24 bitlines low, so the total static power is 130µW • (31 + 24) = 7.2 mW. standby static currents in the 10’s to 100’s of µA. Eventually, static power dissipation may become comparable to dynamic power even for high-power systems. Of course, static dissipation can occur in gates such as pseudo-nMOS gates where there is a direct path between power and ground. If such gates are used, this contention current must be factored into the total static power dissipation of the chip. 4.4.2 Dynamic Dissipation The primary dynamic dissipation component is charging the load capacitance. Suppose a load C is switched between GND and VDD at an average frequency of fsw. Over any given interval of time T, the load will be charged and discharged Tfsw times. Current flows from VDD to the load to charge it. Current then flows from the load to GND during discharge. In one complete charge/discharge cycle, a total charge of Q = CVDD is thus transferred from VDD to GND. The average dynamic power dissipation is T Pdynamic = V 1 i DD (t )V DD dt = DD T∫ T 0 T ∫ iDD (t )dt (4.28) 0 Taking the integral of the current over some interval T as the total charge delivered during that time, we simplify to Pdynamic = V DD [Tf swCV DD ] = CV DD 2 f sw T (4.29) Weste04r4.fm Page 101 Monday, January 5, 2004 1:24 AM 4.4 POWER DISSIPATION Example Our 200M transistor digital system from the example on page 99 uses static CMOS for the logic gates with an average activity factor of 0.1 The memory arrays are divided into banks and only the necessary bank is activated so the effective memory activity factor is 0.05. Assume transistors have a gate capacitance of about 2 fF/µm. Neglecting wire capacitance, estimate the dynamic power consumption per MHz of the system. Solution: There are (20 • 106 logic transistors) • (12 λ • (0.05 µm/λ) • (2 fF/µm) = 24 nF of logic transistors and (180 • 106 memory transistors) • (4 λ) • (0.05 µm/λ) = 72 nF of memory transistors. The power consumption is [(0.1) • (24 • 10–9) + (0.05) • (72 • 10–9) ] • (1.2)2 • f = 8.6 mW/MHz, or 8.6 W at 1 GHz. Because most gates do not switch every clock cycle, it is often more convenient to express switching frequency f sw as an activity factor α times the clock frequency f. Now the dynamic power dissipation may be rewritten as: 2 Pdynamic = αCV DD f (4.30) A clock has an activity factor of α = 1 because it rises and falls every cycle. Most data has a maximum activity factor of 0.5 because it transitions only once each cycle. Static CMOS logic has been empirically determined to have activity factors closer to 0.1 because some gates maintain one output state more often than another and because real data inputs to some portions of a system often remain constant from one cycle to the next. Because the input rise/fall time is greater than zero, both nMOS and pMOS transistors will be ON for a short period of time while the input is between Vtn and VDD – |Vtp|. This results in an additional “short circuit” current pulse from VDD to GND and typically increases power dissipation by about 10% [Veendrick84]. Short circuit power dissipation occurs as both pullup and pulldown networks are partially ON while the input switches [Veendrick84]. It increases as edge rates become slower because both networks are ON for more time. However, it decreases as load capacitance increases because with large loads the output only switches a small amount during the input transition, leading to a small Vds across one of the transistors. Unless the input edge rate is much slower than the output edge rate, short circuit current is a small fraction (< 10%) of current to the load and can be ignored in hand calculations. It is good to use relatively crisp edge rates at the inputs to gates with wide transistors to minimize their short circuit current. 4.4.3 Low-power Design Total power dissipation is the sum of the static and dynamic dissipation components. Dynamic dissipation has historically been far greater than static power when systems are active, and hence, static power is often ignored, although this will change as gate and sub- 101 Weste04r4.fm Page 102 Monday, January 5, 2004 1:24 AM 102 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION threshold leakage increase. Many tools are available to assist with power estimation; these are discussed further in Sections 5.5.4 and 8.4.1.7. Power dissipation has become extremely important to VLSI designers. For high-performance systems such as workstations and servers, dynamic power consumption per chip is often limited to about 150 W by the amount of heat that can be managed with aircooled systems and cost-effective heatsinks. This number increases slowly with advances in heatsink technology and can be increased significantly with expensive liquid cooling, but has not kept pace with the growing power demands of systems. Therefore, performance may be limited by the inability to cool huge systems with power-hungry circuits operating at high speeds. For battery-based systems such as laptops, cell phones, and PDAs, power consumption sets the battery life of the product. In these systems, most or all of the switching activity may be stopped in an idle or “sleep” mode. Hence, in addition to dynamic power while active, static power consumption may limit the battery life while idle. Many papers and books have been written on low-power design. Unfortunately, there are no “silver bullets”; low-power consumption is generally achieved by careful design. Power reduction techniques can be divided into those that reduce dynamic power and those that reduce static power. 4.4.3.1 Dynamic Power Reduction If a process is selected with sufficiently high threshold voltages and oxide thicknesses, static dissipation is small and dynamic dissipation usually dominates while the chip is active. EQ (4.30) shows that dynamic power is reduced by decreasing the activity factors, the switching capacitance, the power supply, or the operating frequency. Activity factor reduction is very important. Static logic has an inherently low activity factor. Clocked nodes such as the clock network and the clock input to registers have an activity factor of 1 and are very power-hungry. Dynamic circuit families, described in Section 6.2.4, have clocked nodes and a high internal activity factor, so they are also costly in power. Clock gating can be used to stop portions of the chip that are idle; for example, a floating point unit can be turned off when executing integer code and a second level cache can be idled if the data is found in the primary cache. A large fraction of power is dissipated by the clock network itself, so entire portions of the clock network can be turned off where possible. The chip can also sense die temperature and cut back activity if the temperature becomes too high. A drawback of activity factor reduction is that if the system transitions rapidly from an idle mode with little switching to a fully active mode, a large di/dt spike will occur. This leads to inductive noise in the power supply network. Some systems throttle execution, limiting the number of functional units that go from idle to active in each cycle. Device-switching capacitance is reduced by choosing small transistors. Minimumsized gates can be used on non-critical paths. Although Logical Effort finds that the best stage effort is about 4, using a larger stage effort increases delay only slightly and greatly reduces transistor sizes. For example, buffers driving I/O pads or long wires may use a stage effort of 8–12 to reduce the buffer size. Interconnect switching capacitance is most Weste04r4.fm Page 103 Monday, January 5, 2004 1:24 AM 4.4 POWER DISSIPATION effectively reduced through careful floorplanning, placing communicating units near each other to reduce wire lengths. Voltage has a quadratic effect on dynamic power. Therefore, choosing a lower power supply significantly reduces power consumption. As many transistors are operating in a velocity-saturated regime, the lower power supply may not reduce performance as much as first-order models predict. Voltage can be adjusted based on operating mode; for example, a laptop processor may operate at high voltage and high speed when plugged into an AC adapter, but at lower voltage and speed when on battery power. If the frequency and voltage scale down in proportion, a cubic reduction in power is achieved. For example, the laptop processor may scale back to 2/3 frequency and voltage to save 70% in power when unplugged. Frequency can also be traded for power. For example, in a digital signal processing system primarily concerned with throughput, two multipliers running at half speed can replace a single multiplier at full speed. At first, this may not appear to be a good idea because it maintains constant power and performance while doubling area. However, if the power supply can also be reduced because the frequency requirement is lowered, overall power consumption goes down. Commonly used metrics in low-power design are power, the power-delay product, and the energy-delay product. Power alone is a questionable metric because it can be reduced simply by computing more slowly. The power-delay (i.e., energy) product is also suspect because the energy can be reduced by computing more slowly at a lower supply voltage. The energy-delay product (i.e., power • delay2) is less prone to such gaming. Overall, the energy-delay product measured in Performance2/Watt (where performance might be in units of SpecInt) normalized for process only varies by about a factor of two across a wide range of general-purpose microprocessor architectures [Gonzalez96]. This suggests that as long as wasteful practices are avoided, there is little you can do to general-purpose processors except trade the energy consumed by a computation against the delay of the computation. The big power gains are to be made not through tweaking of circuits but by reconsidering algorithms. For example, the Fast Fourier Transform requires far fewer arithmetic operations and hence less power than a Discrete Fourier Transform. Signal-processing systems using datapaths hardwired to a particular operation consume far less power than general-purpose processors delivering the same performance because the datapaths eliminate unnecessary control units. 4.4.3.2 Static Power Reduction Static power reduction involves minimizing Istatic. Some circuit techniques such as analog current sources and pseudo-nMOS gates intentionally draw static power. They can be turned off when they are not needed. Recall that the subthreshold leakage current for Vgs < Vt is V gs −V t I ds = I ds0e nv T −V ds 1 − e v T (4.31) 103 Weste04r4.fm Page 104 Monday, January 5, 2004 1:24 AM 104 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION V t = V t 0 − ηV ds + γ ( φ s + V sb − φ s ) (4.32) where the η term describes drain-induced barrier lowering and the γ term describes the body effect. For any appreciable Vds, the term in brackets approaches unity and can be discarded. The remaining term can be reduced by increasing the threshold voltage Vt0, reducing Vgs, reducing Vds, increasing Vsb, or lowering the temperature. Subthreshold leakage power is already a major problem for battery-powered designs in the 180 nm generation and will be growing exponentially as power supplies and threshold voltages are scaled down in future processes. Many low-power systems need high performance while active and low leakage while idle. The high-performance requirement entails relatively low thresholds, which contribute excessive leakage current in the idle mode. As mentioned in Section 4.4.1, selective application of multiple threshold voltages can maintain performance on critical paths with low-Vt transistors while reducing leakage on other paths with high-Vt transistors. In low-power battery-operated devices, leakage specifications may be given at 40° C rather than 110° C because battery life is most important in the range of normal ambient temperatures. Another way to control leakage is through the body voltage using the body effect. For example, low-Vt devices can be used and a reverse body bias (RBB) can be applied during idle mode to reduce leakage [Kuroda96]. Alternatively, higher-Vt devices can be used, and then a forward body bias (FBB) can be applied during active mode to increase performance [Narendra03]. As we will see in Section 4.7.3, threshold voltages vary from one die to another on account of manufacturing variations. An adaptive body bias (ABB) can compensate and achieve more uniform transistor performance despite the variations [Narendra99, Tschanz02]. In any case, the body bias should be kept to less than about 0.5 V. Too much reverse body bias leads to greater junction leakage through a mechanism called band-toband tunneling [Keshavarzi01], while too much forward body bias leads to substantial current through the body to source diodes. Applying a body bias requires additional power supply rails to distribute the substrate and well voltages. For example, a RBB scheme for a 1.8 V n-well process could bias the ptype substrate at VBBn = –0.4 V and the n-well at VBBp = 2.2 V. Figure 4.27 shows a schematic and cross-section of an inverter using body bias. In an n-well process, all nMOS transistors share the same p substrate and must use the same VBBn. In a triple-well process, groups of transistors can use different p-wells isolated from the substrate and thus can use different body biases. The well and substrate carry little current, so the bias voltages are relatively easy to generate and distribute. Alternatively, the source voltage can be raised in sleep mode. This has the double benefit of reducing Vds (to increase Vt through reduced DIBL and to also reduce gate leakage) as well as increasing Vsb (to increase Vt through the body effect). However, the source does carry significant current, so generating a stable and adjustable source voltage rail is challenging. Weste04r4.fm Page 105 Monday, January 5, 2004 1:24 AM 4.4 POWER DISSIPATION VBBp VDD A (a ) Y GND VBBn A VBBn GND VDD Y p+ n+ n+ p+ p+ VBBp n+ n-well p-substrate (b) Substrate Tap Well Tap FIG 4.27 Body bias Reducing VDD in standby mode reduces the drain-induced barrier lowering contribution to leakage. It also decreases gate leakage in processes where that component is important. The supply should be maintained at a high enough level to preserve the state of the system [Clark02]. Yet another method of reducing idle leakage current in low-power systems is to turn off the power supply entirely. This could be done externally with the voltage regulator or internally with a series transistor. Multiple Threshold CMOS circuits (MTCMOS) use lowVt transistors for computation and a high-Vt transistor as a switch to disconnect the power supply during idle mode, as shown in Figure 4.28 [Mutoh95]4. The high-Vt device is connected between the true VDD and the virtual VDDV rails connected to the logic gates. The extra transistor increases the impedance between the true and virtual power supply, causing greater power supply noise and gate delay. Bypass capacitance between VDDV and GND stabilizes the supply somewhat, but the capacitance is discharged each time VDDV is disconnected, contributing to the power consumption. Even using a very wide high-Vt transistor, MTCMOS is only suited to systems with small power demands. The pMOS body should be tied to V DD so both V DD and V DDV lines must be routed to all cells. MTCMOS uses carefully designed registers connected to the true supply rails to retain state during idle mode [Kao01]. 4 Do not confuse MTCMOS with the use of high-Vt transistors in noncritical gates and low-Vt transistors used in critical gates described earlier. 105 Weste04r4.fm Page 106 Monday, January 5, 2004 1:24 AM 106 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION VDD Standby VDDV High-Vt Sleep Transistor Low-Vt Logic Transistors Bypass Capacitance GND FIG 4.28 MTCMOS VDD VDD I1 I2 The leakage through two series OFF transistors is much lower than that of a single transistor because of the stack effect [Ye98, Narendra01]. In Figure 4.29(a), the single transistor has a relatively low threshold because of draininduced barrier lowering from the high drain voltage. In Figure 4.29(b), node x rises to about 100 mV. The threshold on the bottom transistor is higher because of the small drain voltage. The top transistor also turns off harder because of the negative Vgs and the body effect. The net result is that I2 may be 10–20 times smaller than I1. Low-power systems can take advantage of this stack effect to put gates with series transistors into a low-leakage sleep mode by applying an input pattern to turn off both transistors. Silicon on Insulator (SOI) circuits are attractive for low-leakage designs because they have a sharper subthreshold current rolloff (smaller n in EQ (4.26)). SOI will be discussed further in Section 6.7. x (a ) 4.5 Interconnect (b) The wires linking transistors together are called interconnect and play a major role in the performance of modern systems. Figure 4.30 shows a pair of adjacent wires. The wires have width w, length l, thickness t, and spacing of s from their neighbors and have a dielectric of height h between them and the conducting layer below. The sum of width and spacing is called the wire pitch. The thickness to width ratio t/w is called the aspect ratio. The dielectric is made of SiO2 or a low-k material. In the early days of VLSI, transistors were relatively slow. Wires were wide and thick and thus had low resistance. Under those circumstances, wires could be treated as ideal equipotential nodes with lumped capacitance. In modern VLSI processes, transistors switch much faster. Meanwhile, wires have become narrower, driving up their resistance to the point that in many paths the wire RC delay exceeds gate delay. Moreover, the wires are packed very closely together and thus a large fraction of their capacitance is to their neighbors. When one wire switches, it tends to affect its neighbor through capacitive coupling; this effect is called crosstalk. On-chip interconnect inductance had been negligible but is now becoming a factor for systems with fast edge rates and closely packed busses. Considering all of these factors, circuit design is now as much about engineering the wires as the transistors that sit underneath. Early CMOS processes had a single metal layer and for many years only two or three layers were available, but with advances in chemical-mechanical polishing it became far more practical to manufacture many metal layers. A 180 nm process typically has about six to eight metal layers and the layer count has been increasing at a rate of about one per generation. Figure 4.31 shows a cross-section of the Intel 180 nm process metal stack [Yang98]. Metal1 wires are thin and built on a tight pitch to provide dense routing within a cell; observe that the pitch is less than 6 λ, while conservative scalable CMOS rules would dictate an 8 λ pitch. Their resistance is high, but that is acceptable because the wires tend to be short. Top-level metal wires are thicker and built on a wide pitch. FIG 4.29 Leakage stack effect Weste04r4.fm Page 107 Monday, January 5, 2004 1:24 AM 4.5 w s l t h FIG 4.30 Interconnect geometry Layer T (nm) W (nm) S (nm) AR 6 1720 860 860 2.0 800 800 2.0 540 540 2.0 320 320 2.2 320 320 2.2 250 250 1.9 1000 5 1600 4 1080 1000 3 2 1 700 700 700 700 700 480 800 Substrate FIG 4.31 Layer stack for 6-metal Intel 180 nm process This low-resistance layer is useful for power, ground, clock, and critical signal routing. In the Intel process, the intermediate layers gradually increase in width and pitch, although in many processes the intermediate layers are uniform. The width and spacing are typically comparable. INTERCONNECT 107 Weste04r4.fm Page 108 Monday, January 5, 2004 1:24 AM 108 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION 4.5.1 Resistance The resistance of a uniform slab of conducting material can be expressed as R= ρl tw (4.33) where ρ is the resistivity5. This expression can be rewritten as R=R l w (4.34) where R? = ρ/t is the sheet resistance and has units of Ω/square. Note that a square is a dimensionless quantity corresponding to a slab of equal length and width. This is convenient because resistivity and thickness are characteristics of the process outside the control of the circuit designer and can be abstracted away into the single sheet resistance parameter. To obtain the resistance of a conductor on a layer, multiply the sheet resistance by the ratio of length to width of the conductor. For example, the resistance of the two shapes in Figure 4.32 are equal because the length-to-width ratio is the same even though the sizes are different. Nonrectangular shapes can be decomposed into simpler regions for which the resistance is calculated [Horowitz83]. w w l w l l t t 1 Rectangular Block R = R (L/W) Ω 4 Rectangular Blocks R = R (2L/2W) Ω = R (L/W) Ω FIG 4.32 Two conductors with equal resistance 5 ρ is used to indicate both resistivity and best stage effort. The meaning should be clear from context. Weste04r4.fm Page 109 Monday, January 5, 2004 1:24 AM 4.5 Table 4.6 shows bulk electrical resisitivities of pure metals [Bakoglu90]. The sheet resistance of thin metal films used in wires tends to be slightly higher, e.g., 2.6 µΩ • cm for Cu and 3.5–4.0 µΩ • cm for Al. Most processes prior to the 0.18 µm generation use aluminum wires. Modern processes often use copper to reduce the resistivity and also to obtain better electromigration characteristics (see Section 4.8.2). Unfortunately, copper must be surrounded by a lower-conductivity diffusion barrier that effectively reduces the wire cross-sectional area and hence raises the resistance. Aluminum does not require such a barrier and thus may actually offer lower resistance for very narrow wires in the future. Electron surface scattering effects in thin conductors also result in somewhat higher resistance for on-chip interconnect than simple bulk resistivity would predict. Table 4.6 Bulk resistivity of pure metals at 22º Metal Silver (Ag) Copper (Cu) Gold (Au) Aluminum (Al) Tungsten (W) Molybdenum (Mo) Titanium (Ti) Resistivity (µΩ • cm) 1.6 1.7 2.2 2.8 5.3 5.3 43.0 Table 4.7 shows typical sheet resistances for the 180 nm process with aluminum interconnect. The upper layers of metal have lower resistivity because they are thicker. Metal resistance is determined by the material (usually Al or Cu). The resistivity of polysilicon, diffusion, and wells is significantly influenced by the doping levels. Polysilicon and diffusion are often silicided with TiSi2 (see Section 3.2.8) to reduce the resistance. Interconnect resistance increases with temperature; this effect is particularly pronounced for wells and diffusion. Contacts and vias also have a resistance associated with them that is dependent on the contacted materials and size of the contact. Typical values are 2–20 Ω. Multiple contacts should be used to form low-resistance connections. Because current crowding tends to occur at the periphery of contacts, design rules dictate multiple small contacts rather than a single large contact, as shown in Figure 4.33. When current turns at a right angle or reverses, a square array of contacts is generally required, while when the flow is in the same direction, fewer contacts can be used. INTERCONNECT 109 Weste04r4.fm Page 110 Monday, January 5, 2004 1:24 AM 110 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Table 4.7 Sheet resistances Sheet Resistance (Ω / ?) Layer Diffusion (silicided) Diffusion (unsilicided) Polysilicon (silicided) Polysilicon (unsilicided) Metal1 Metal2 Metal3 Metal4 Metal5 Metal6 3-10 50-200 3-10 50-400 0.08 0.05 0.05 0.03 0.02 0.02 FIG 4.33 Multiple vias for low-resistance connections 4.5.2 Capacitance An isolated wire over the substrate can be modeled as a conductor over a ground plane. The wire capacitance has two major components: the parallel plate capacitance of the bottom of the wire to ground and the fringing capacitance arising from fringing fields along the edge of a conductor with finite thickness. In addition, a wire adjacent to a second wire on the same layer can exhibit capacitance to that neighbor. These effects are illustrated in Figure 4.34. The classic parallel plate capacitance formula is C= ε ox wl h (4.35) Note that oxides are often doped with phosphorous to trap ions before they damage transistors; this oxide has εox ≈ 4.1ε0, as compared to 3.9ε0 for an ideal oxide or lower for lowk dielectrics. The fringing capacitance is more complicated to compute and requires a numerical field solver for exact results. A number of authors have proposed approximations to this Weste04r4.fm Page 111 Monday, January 5, 2004 1:24 AM 4.5 111 INTERCONNECT calculation [Barke88, Ruehli73, Yuan82]. One intuitively appealing approximation treats a lone conductor above a ground plane as a rectangular middle section with two hemispherical end caps, as shown in Figure 4.35 [Yuan82]. The total capacitance is assumed to be the sum of a parallel plate capacitor of width w – t/2 and a cylindrical capacitor of radius t/2. This results in an expression for the capacitance that is accurate within 10% for aspect ratios less than 2 and t ≈ h. w − C = ε ox l h t 2 2π + 2h 2h 2h + ln1 + + 2 t tt w (4.36) s t h FIG 4.34 Effect of fringing fields on capacitance An empirical formula that is computationally efficient and relatively accurate is [Meijs84, Barke88] 0.25 0.5 w w t C = ε ox l + 0.77 + 1.06 + 1.06 h h h good to 6% for aspect ratios less than 3.3. These formulae do not account for neighbors on the same layer or higher layers. Capacitance interactions between layers can become quite complex in modern multilayer CMOS processes. A conservative upper bound on capacitance can be obtained assuming that the layers above and below the conductor of interest are solid ground planes. Similarly, a lower bound can be obtained assuming there are no other conductors in the system except the substrate. The upper bound can be used for propagation delay and power estimation while the lower bound can be used for contamination delay calculations before layout information is available. A (4.37) Half cylinders w t Parallel plate h FIG 4.35 Yuan & Trick capacitance model including fringing fields Weste04r4.fm Page 112 Monday, January 5, 2004 1:24 AM 112 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION cross-section of the model used for capacitance upper bound calculations is shown in Figure 4.36. The total capacitance of the conductor of interest is the sum of its capacitance to the layer above, the layer below, and the two adjacent conductors. If the layers above and below are not switching6, they can be modeled as ground planes and this component of capacitance is called Cgnd. Theoretically, wires will have some capacitance to further neighbors, but in practice this capacitance is normally small enough to ignore because most electric fields terminate on the nearest conductors. C gnd = C bot + C top C total = C gnd + 2C adj s (4.38) w Layer n + 1 h2 Ctop t h1 Layer n Cbot Cadj Layer n – 1 FIG 4.36 Multilayer capacitance model The capacitances can be computed by generating a table of data with a field solver such as FastCap [Nabors92] or HSPICE. Tables can be generated for different widths and spacings on each layer and you can interpolate between entries if necessary. For example, Table 4.8 tabulates capacitance for the layer stack from Figure 4.31 using a SiOF low-k dielectric with εox = 3.55ε0 and assuming solid planes of metal above and below. The width w and spacing s indicate multiples of the minimum allowable for the particular metal layer. Three columns for each metal layer indicate the capacitance to each adjacent neighbor, to ground (i.e., the planes above and below, which are assumed on average to not be moving), and total (EQ (4.38)). The capacitance to neighbors accounts for more than 50% of the capacitance for narrow wires. We will see in Section 4.5.4 that coupling to these neighbors has a significant effect on delay and signal integrity and limits the acceptable aspect ratios. The capacitance for metal2 and metal3 lines should be equal because they have the same geometry; the minor variations reflect finite numerical precision in the HSPICE field solver. Table 4.9 shows the same information assuming no metal on other planes. The overall capacitance is slightly smaller because there are no nearby planes for fields to terminate on, but the capacitance to the adjacent neighbors is higher because many of the fringing fields now terminate on the neighbors instead. 6 Or at least consist of a large number of orthogonal conductors that on average cancel each other’s switching activities. Weste04r4.fm Page 113 Monday, January 5, 2004 1:24 AM 4.5 INTERCONNECT 113 Table 4.8 Capacitance table for 180 nm process (aF/µm) with metal planes above and below w 1 1 1 1 1.5 1.5 1.5 1.5 2 2 2 2 3 3 3 3 4 4 4 4 6 6 6 6 s 1 1.5 2 ∞ 1 1.5 2 ∞ 1 1.5 2 ∞ 1 1.5 2 ∞ 1 1.5 2 ∞ 1 1.5 2 ∞ Metal1 Cadj 84 57 42 0 85 58 43 0 86 59 44 0 87 60 44 0 87 60 44 0 87 59 44 0 Cgnd 43 52 60 112 53 62 70 123 63 72 80 134 83 92 101 155 104 113 122 176 146 155 163 218 Metal2 Metal3 Metal4 Metal5 Metal6 Ctot Cadj Cgnd Ctot Cadj Cgnd Ctot Cadj Cgnd Ctot Cadj Cgnd Ctot Cadj Cgnd Ctot 210 88 57 232 88 57 232 77 78 232 73 94 240 82 64 227 166 58 68 184 58 68 184 49 93 191 45 112 202 54 76 184 144 42 78 162 42 78 162 33 106 173 30 126 186 39 85 164 112 0 132 132 0 132 132 0 154 154 0 167 167 0 137 137 224 89 70 248 89 70 248 77 99 253 73 119 266 83 80 246 178 59 82 199 59 82 199 49 114 212 45 137 227 56 91 202 156 42 92 177 42 92 177 34 127 194 30 151 211 41 101 182 123 0 147 147 0 147 147 0 174 174 0 192 192 0 154 154 236 89 84 263 89 84 263 77 119 274 73 144 291 84 95 264 190 59 96 214 59 96 214 49 134 232 45 162 252 57 107 220 167 43 106 191 43 106 191 34 147 215 30 176 236 41 117 199 134 0 162 162 0 162 162 0 195 195 0 217 217 0 171 171 258 90 113 292 90 113 292 77 160 315 73 194 341 86 127 298 211 59 124 243 59 124 243 49 175 273 45 212 302 58 138 254 189 43 135 220 43 135 220 34 188 256 30 226 286 42 148 233 155 0 191 191 0 191 191 0 236 236 0 269 269 0 204 204 279 90 141 320 90 141 320 77 201 355 73 245 390 86 159 331 232 59 153 272 59 153 272 49 216 314 45 262 352 58 170 287 210 43 164 249 43 164 249 34 229 296 30 276 336 43 180 266 176 0 219 219 0 219 219 0 278 278 0 321 321 0 237 237 320 89 199 377 89 199 377 77 284 437 72 345 490 86 223 396 274 59 210 328 59 210 328 49 299 396 45 362 451 58 235 351 252 42 221 306 42 221 306 33 312 378 30 376 436 43 245 331 218 0 276 276 0 276 276 0 360 360 0 422 422 0 301 301 Figure 4.37 plots total capacitance of a metal2 line from the same process as a function of width for various spacings. For an isolated wire above the substrate, the capacitance is strongly influenced by spacing between conductors. For a wire sandwiched between metal1 and metal3 planes, the capacitance is higher and is more sensitive to the width (determining parallel plate capacitance) but less sensitive to spacing once the spacing is significantly greater than the wire thickness. In either case, the y-intercept is greater than zero so doubling the width of a wire results in less than double the total capacitance. Tight-pitch metal lines have a capacitance of roughly 0.2 fF/µm. Weste04r4.fm Page 114 Monday, January 5, 2004 1:24 AM 114 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Table 4.9 Capacitance table for 180 nm process (aF/µm) with substrate below and nothing above w 1 1 1 1 1.5 1.5 1.5 1.5 2 2 2 2 3 3 3 3 4 4 4 4 6 6 6 6 s 1 1.5 2 ∞ 1 1.5 2 ∞ 1 1.5 2 ∞ 1 1.5 2 ∞ 1 1.5 2 ∞ 1 1.5 2 ∞ Metal1 Cadj 89 63 49 0 92 66 52 0 94 68 53 0 98 71 56 0 100 73 58 0 102 76 61 0 Cgnd 28 33 37 87 33 38 42 94 38 43 47 100 48 52 57 112 58 62 67 124 77 82 86 147 Metal2 Ctot 206 159 136 87 217 170 145 94 227 178 154 100 243 194 169 112 257 208 182 124 282 233 207 147 Metal3 Metal4 Metal5 Metal6 Cadj Cgnd Ctot Cadj Cgnd Ctot Cadj Cgnd Ctot Cadj Cgnd Ctot Cadj Cgnd Ctot 101 19 220 96 16 208 102 15 219 96 15 208 97 14 207 73 22 167 70 18 158 75 17 166 70 17 158 71 15 158 58 24 140 57 20 133 60 18 138 57 19 132 58 17 132 0 70 71 0 60 61 0 59 60 0 60 60 0 55 55 104 21 230 100 17 218 106 16 229 100 17 217 101 15 217 76 24 176 74 19 167 78 18 174 74 19 167 75 16 166 61 27 148 60 21 140 63 20 146 60 21 140 61 18 139 0 75 75 0 64 64 0 62 62 0 63 63 0 58 58 107 24 238 103 19 225 110 18 237 103 18 225 104 16 224 78 26 183 76 21 174 81 19 181 77 20 173 78 17 173 63 29 154 62 22 147 66 21 152 62 22 146 63 19 145 0 78 79 0 67 67 0 65 65 0 66 66 0 60 60 111 29 251 108 22 237 114 20 249 108 21 237 109 18 236 82 31 195 81 24 185 85 22 192 81 23 184 82 20 183 66 34 166 66 25 157 69 24 162 66 25 156 67 21 155 0 86 86 0 72 72 0 69 70 0 71 72 0 64 65 114 33 262 111 25 247 118 23 258 111 24 246 112 20 245 85 36 205 83 27 194 88 25 201 83 26 193 85 22 192 68 38 175 68 28 165 72 26 171 68 28 165 70 23 163 0 92 93 0 77 77 0 74 75 0 76 76 0 68 69 118 43 279 115 31 262 122 28 272 115 31 260 116 25 258 88 46 222 87 33 208 92 30 215 87 33 207 89 26 204 72 48 191 72 35 179 76 31 184 72 34 178 73 28 175 0 105 105 0 86 86 0 82 83 0 85 85 0 75 76 In practice, the layers above and below the conductor of interest are neither solid planes nor totally empty. One can extract capacitance more accurately by interpolating between these two extremes based on the density of metal on each level. [Chern92] gives formulae for this interpolation accurate to within 10%. However, if the wiring above and below is fairly dense (e.g., a bus on minimum pitch), it is well-approximated as a plane. Dense wire fill is added to many chips for mechanical stability and etch uniformity, making this approximation even more appropriate. Weste04r4.fm Page 115 Monday, January 5, 2004 2:03 AM 4.5 115 INTERCONNECT 400 350 300 M1, M3 planes s = 320 s = 480 s = 640 s= 200 8 Ctotal ( aF/µm) 250 Isolated s = 320 150 s = 480 s= 8 s = 640 100 50 0 0 500 1000 1500 2000 w (nm) FIG 4.37 Capacitance of metal2 line as a function of width and spacing 4.5.3 Delay Interconnect increases circuit delay for two reasons. First, the wire capacitance adds loading to each gate. Second, long wires have significant resistance that contributes distributed RC delay or flight time. It is straightforward to add wire capacitance to the Elmore delay calculations of Section 4.2.1, so in this section we focus on the RC delay. The distributed resistance and capacitance of a wire can be approximated with a number of lumped elements. Three standard approximations are the L-model, π -model, and T-model, sonamed because of their shape. Figure 4.38 shows how a distributed RC circuit is equivalent to N distributed RC segments of proportionally less resistance and capacitance, and how these segments N segments R/N R/N R/N R C C/N C/N R R C C/2 L-mode l R/N C/N C/N R/2 R/2 C/2 -mode l C T-mode l FIG 4.38 Lumped approximation to distributed RC circuit Weste04r4.fm Page 116 Monday, January 5, 2004 2:03 AM 116 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Example Consider a 5 mm long, 0.32 µm wide metal2 wire in a 0.18 µm process. The sheet resistance is 0.05 Ω/? and the capacitance is 0.2 fF/µm. Construct a 3-segment π-model for the wire. Solution: The wire is 5000 µm/0.32 µm = 15625 squares in length. The total resistance is (0.05 Ω/?) • (15625 ?) = 781 Ω. The total capacitance is (0.2 fF/ µm) • (5000 µm) = 1 pF. Each π-segment has one-third of this resistance and capacitance. The π -model is shown in Figure 4.39(a); adjacent capacitors can be merged as shown in Figure 4.39(b). 260 2 60 (a ) 1 67 fF 167 f F 1 67 fF 167 f F 1 67 fF 167 f F 260 2 60 333 fF 333 fF 260 (b) 1 67 fF 260 167 f F FIG 4.39 3-segment π-model for wire Example A 10x unit-sized inverter drives a 2x inverter at the end of the 5 mm wire from the previous example. The gate capacitance is C = 2 fF/µm and the effective resistance is R = 2.5 kΩ • µm for nMOS transistors. Estimate the propagation delay using the Elmore delay model; neglect diffusion capacitance. Solution: A unit inverter has a 4 λ = 0.36 µm wide nMOS transistor and an 8 λ = 0.72 µm wide pMOS transistor. Hence, the unit inverter has an effective resistance of (2.5 kΩ • µm)/(0.36 µm) = 6.9 kΩ and a gate capacitance of (0.36 µm + 0.72 µm) • (2 fF/µm) = 2 fF. Larger inverters have proportionally more capacitance and less 7 81 resistance. Figure 4.40 shows an equivalent circuit for the system using a single- 690 5 00 fF 500 fF 4 fF segment π-model. The Elmore delay is tpd = (690 Ω) • (500 fF) + (690 Ω + 781 Ω) • Driver Wire Load (500 fF + 4 fF) = 1.1 ns. The capacitance FIG 4.40 Equivalent circuit for example of the long wire dominates the delay; the capacitance of the 2x inverter is negligible in comparison. Weste04r4.fm Page 117 Monday, January 5, 2004 1:24 AM 4.5 INTERCONNECT 117 can be modeled with lumped elements. As the number of segments approaches infinity, the lumped approximation will converge with the true distributed circuit. The L-model is a very poor choice because a large number of segments are required for accurate results. The π-model is much better; three segments are sufficient to give results accurate to 3% [Sakurai83]. The T-model is comparable to the π-model, but produces circuits with more nodes that are slower to solve by hand or with a circuit simulator. Therefore, it is common practice to model long wires with a 3- or 4-segment π-model for simulation. These results make sense because the Elmore delay of a single-segment L-model is RC while the Elmore delay of a single-segment π- or T-model is RC/2. Single-segment πmodels are a reasonable approximation for hand calculations. Because both wire resistance and wire capacitance increase with length, wire delay grows quadratically with length. Using thicker and wider wires, lower-resistance metals such as copper, and lower-dielectric constant insulators helps, but long wires nevertheless have unacceptable delay. Section 4.6.4 describes how repeaters can be used to break a long wire into multiple segments such that the overall delay becomes a linear function of length. Polysilicon and diffusion wires (sometimes called runners) have high resistance, even if silicided. Diffusion also has very high capacitance. Therefore, it is good practice to avoid running diffusion any further than absolutely necessary and to only use polysilicon for wiring within a cell. 4.5.4 Crosstalk As reviewed in Figure 4.41, wires have capacitance to their adjacent neighbors as well as to ground. When wire A switches, it tends to bring its neighbor B along with it on account of capacitive coupling, also called crosstalk. If B is supposed to switch simultaneously, this may increase or decrease the switching delay. If B is not supposed to switch, crosstalk causes noise on B. We will see that the impact of crosstalk depends on the ratio of Cadj to the total capacitance. Note that the load capacitance is included in the total, so for short wires and large loads, the load capacitance dominates and crosstalk is unimportant. Conversely, for long wires crosstalk is very important. A Cgnd B Cadj Cgnd FIG 4.41 Capacitances 4.5.4.1 Crosstalk Delay Effects If both a wire and its neighbor are switching, the direction of the switching affects the amount of charge that must be delivered and the delay of the switching. Table 4.10 summarizes this effect. The charge delivered to the coupling capacitor is Q = Cadj∆V, where ∆V is the change in voltage between A and B. If A switches but B does not, ∆V = VDD. The total capacitance effectively seen by A is just the capacitance to ground and to B. If both A and B switch in the same direction, ∆V = 0. Hence, no charge is required and Cadj is effectively absent for delay purposes. If A and B switch in the opposite direction, ∆V = 2VDD. Twice as much charge is required. Equivalently, the capacitor can be treated as being effectively twice as large switching through VDD. This is analogous to the Miller effect discussed in Section 4.2.5.4. The Miller Coupling Factor (MCF) to adjacent neighbor and to ground Weste04r4.fm Page 118 Monday, January 5, 2004 1:24 AM 118 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION describes how the capacitance to adjacent wires is multiplied to find the effective capacitance. Some designers use MCF = 1.5 as a statistical compromise when estimating propagation delays before layout information is available. Table 4.10 Dependence of effective capacitance on switching direction B ∆V Ceff(A) constant VDD Cgnd + Cadj 1 switching same direction as A 0 Cgnd 0 switching opposite A 2VDD Cgnd + 2Cadj 2 MCF Example Each wire in a pair of 1 mm lines has capacitance of 0.1 fF/µm to ground and 0.1 fF/µm to its neighbor. Each line is driven by an inverter with a 1 kΩ effective resistance. Estimate the contamination and propagation delays of the path. Neglect parasitic capacitance of the inverter and resistance of the wires. Solution: We find Cgnd = Cadj = (0.1 fF/µm) • (1000 µm) = 0.1 pF. The delay is RCeff. The contamination delay is the minimum possible delay, which occurs when both wires switch in the same direction. In that case, Ceff = Cgnd and the delay is tcd = (1kΩ) • (0.1 pF) = 100 ps. The propagation delay is the maximum possible delay, which occurs when both wires switch in opposite directions. In this case, Ceff = Cgnd + 2Cadj and the delay is tpd = (1kΩ) • (0.3 pF) = 300 ps. A conservative design methodology assumes neighbors are switching when computing propagation and contamination delays (MCF = 2 and 0, respectively). This leads to a wide variation in the delay of wires. A more aggressive methodology tracks the time window during which each signal can switch. Thus, switching neighbors must be accounted for only if the potential switching windows overlap. Similarly, the direction of switching can be considered. For example, dynamic gates described in Section 6.2.4 precharge high and then fall low during evaluation. Thus, a dynamic bus will never see opposite switching during evaluation. 4.5.4.2 Crosstalk Noise Effects Suppose wire A switches while B is supposed to remain constant. This introduces noise as B partially switches. We call A the aggressor or perpetrator and B the victim. If the victim is floating, we can model the circuit as a capaci- Weste04r4.fm Page 119 Monday, January 5, 2004 1:58 AM 4.5 INTERCONNECT 119 tive voltage divider to compute the victim noise, as shown in Figure 4.42. ∆Vaggressor is normally VDD. ∆V victim = C adj C gnd − v + C adj ∆V aggressor (4.39) If the victim is actively driven, the driver will supply current to oppose and reduce the victim noise. We model the drivers as resistors, as shown in Figure 4.43. The peak noise becomes dependent on the time constant ratio k of the aggressor to the victim [Ho01]: ∆V victim = C adj 1 ∆V aggressor C gnd − v + C adj 1 + k (4.40) where k= τaggressor τ victim = ( Raggressor C gnd − a + C adj ( R victim C gnd − v + C adj ) Figure 4.44 shows simulations of coupling when the aggressor is driven with a unit inverter; the victim is undriven or driven with an inverter of half, equal, or twice the size of the aggressor; and Cadj = Cgnd . Observe that when the victim is floating, the noise remains indefinitely. When the victim is driven, the driver restores the victim. Larger (faster) drivers oppose the coupling sooner and result in noise that is a smaller percentage of the supply voltage. Note that during the noise event the victim transistor is in its linear region while the aggressor is in saturation. For equal-sized drivers, this means Raggressor is two to four times Rvictim, with greater ratios arising from more velocity saturation [Ho01]. In general, EQ (4.40) is conservative, especially when wire resistance is included [Vittal99]. It is often used to flag nets where coupling can be a problem; then simulations can be performed to calculate the exact coupling noise. Coupling noise is of greatest importance on weakly driven nodes where k < 1. We have only considered the case of a single neighbor switching. When both neighbors switch, the noise will be twice as great. We have also modeled the layers above and below as AC ground planes, but wires on these layers are likely to be switching. For a long line, you can expect about as many ) (4.41) Aggressor Vaggressor Cadj Victim Cgnd-v Vvictim FIG 4.42 Coupling to floating victim Raggressor Aggressor Cgnd-a Vaggressor Cadj Rvictim Victim Cgnd-v FIG 4.43 Coupling to driven victim Vvictim Weste04r4.fm Page 120 Monday, January 5, 2004 2:05 AM 120 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Aggressor 1.8 1.5 1.2 Victim (undriven): 50% 0.9 0.6 Victim (half size driver): 16% Victim (equal size driver): 8% 0.3 Victim (double size driver): 4% 0 0 200 400 600 800 1000 1200 1400 1800 2000 t (ps) FIG 4.44 Waveforms of coupling noise lines switching up and switching down, giving no net contribution to delay or noise. However, a short line running over a 64-bit bus in which all 64 bits are simultaneously switching from 0 to 1 will be strongly influenced by this switching. 4.5.5 Inductance Most design tools consider only interconnect resistance and capacitance. Inductance is beginning to be important for accurately modeling on-chip power grids, clock networks, and wide busses. While the industry still has a very limited understanding of inductive effects, some of the key issues are introduced in this section. Although we generally discuss current flowing from a gate output to charge or discharge a load capacitance, current really flows in loops. The return path for a current loop is usually the power or ground network; at the frequencies of interest, the power supply is an “AC ground” because the bypass capacitance forms a low-impedance path between VDD and GND. Currents flowing around a loop generate a magnetic field proportional to the area of the loop and the amount of current. Changing the current requires supplying energy to change the magnetic field. This means that changing currents induce a voltage proportional to the rate of change. The constant of proportionality is called the inductance, L7. 7 L is used to indicate both inductance and transistor channel length. The meaning should be clear from context. Weste04r4.fm Page 121 Monday, January 5, 2004 1:24 AM 4.5 V =L dI dt (4.42) Inductance and capacitance also set the speed of light in a medium. Even if the resistance of a wire is zero leading to zero RC delay, the speed of light flight-time along a wire of length with inductance and capacitance per unit length of L and C is t pd = l LC (4.43) If the current return paths are the same as the conductors on which electric field lines terminate, the signal velocity v is v= 1 LC = 1 ε ox µ 0 = c 3.9 (4.44) where µ0 is the magnetic permeability of free space (4π • 10–7 H/m) and c is the speed of light in free space (3 • 108 m/s). In other words, signals travel about half the speed of light. Using low-k (< 3.9) dielectrics raises this velocity. However, many signals have electric fields terminating on nearby neighbors, but currents returning in more distant power supply lines. This raises the inductance and reduces the signal velocity. Changing magnetic fields in turn produce currents in other loops. This means that signals on one wire can inductively couple onto another; this is called inductive crosstalk. The inductance of a conductor of length and width w located a height h above a ground plane is approximately L=l µ 0 8h w ln + 2π w 4 h (4.45) assuming w < h and thickness is negligible. Typical on-chip inductance values are in the range of 0.15–1.5 pH/µm depending on the proximity of the power or ground lines. (Wires near their return path have smaller current loops and lower inductance.) Current flows along the path of lowest impedance Z = R + jωL. At high-frequency ω, impedance becomes dominated by inductance. The inductance is minimized if the current flows only near the surface of the conductor closest to the return path. This skin effect can reduce the effective cross-sectional area of thick conductors and raise the effective resistance at high frequency. The skin depth for a conductor is δ= 2ρ ωµ 0 (4.46) INTERCONNECT 121 Weste04r4.fm Page 122 Monday, January 5, 2004 1:24 AM 122 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Example Find the skin depth for signals with 100 ps edge rates on copper interconnect (ρ = 1.7 • 10–8 Ω • m). Solution: 100 ps edge rates correspond to energy at 1.67 GHz. δ= ( 2π ( 2 1.7 × 10 −8 Ω • m • 1.67 )( ) × 109 rad / s 4 π × 10 −7 H / m ) = 1.6µm (4.47) Skin depth is not a major consideration at the time of writing, but will limit the benefit of thick wires at higher frequencies in the future. The frequency of importance is the highest frequency with significant power in the Fourier transform of the signal. This is not the chip-operating frequency, but rather is associated with the faster edges. It can be approximated as ω= 2π 6t rf (4.48) Extracting inductance in general is a three-dimensional problem and is extremely time-consuming for complex geometries. Inductance depends on the entire loop and therefore cannot be simply decomposed into sections in the way capacitance is. It is therefore impractical to extract the inductance from a chip layout. Instead, usually inductance is extracted using tools such as FastHenry [Kamon94] for simple test structures intended to capture the worst cases on the chip. This is only possible when the power supply network is highly regular. Power planes are ideal but require a large amount of metal resources. Dense power grids are usually the preferred alternative. Gaps in the power grid force current to flow around the gap, increasing the loop area and greatly increasing inductance. Moreover, large loops couple magnetic fields through other loops formed by conductors at a distance. Therefore, mutual inductive coupling can occur over a long distance, especially when the return path is far from the conductor. Incorporating inductance into simulations is also difficult. Instead, designers usually generate design rules that allow inductance to generally be ignored as long as the rules are followed. Inductance has always been important for integrated circuit packages where the physical dimensions are large, as will be discussed in Section 12.2.3. On-chip inductance is important for wires where the speed of light flight time is longer than either the rise times of the circuits or the RC delay of the wire. Because speed of light flight time increases linearly and RC delay increases quadratically with length, we can estimate the set of wire lengths for which inductance is relevant [Ismail99]. Weste04r4.fm Page 123 Monday, January 5, 2004 1:59 AM 4.5 Example Consider a metal2 signal line with a sheet resistance of 0.05 Ω/? and a width of 0.5 µm. The capacitance is 0.2 fF/µm and inductance is 0.5 pH/µm. Compute the velocity of signals on the line and plot the range of lengths over which inductance matters as a function of the rise time. Solution: The velocity is v= 1 LC = 1 (0.5 pH / µm)(0.2 fF / µm) Note that this is 100 mm/ns or 1 mm / 10 ps. The resistance is (0.1 Ω/?) • (1?/0.5 µm) = 0.2 Ω/ µm. Figure 4.45 plots the length of wires for which inductance is rele vant against r ise times. Above the horizontal line, wires greater than 500 µm are limited by RC delay rather than LC delay. To the right of the diagonal line, rise times are greater than the LC delay. Only in the region between these lines is inductance relevant to delay calculations. This region has very fast edge rates, so inductance is not very important to the delay of highly resistive signal lines at the time of this writing. = 108 m / s = 1 c 3 (4.49) Wire Length 10 mm RC Delay Dominates Rise Time and RC Dominate 1 mm Inductance Matters 100 m Rise Time Dominates 10 m tr 1 ps 10 ps 100 ps 1 ns FIG 4.45 Wire lengths and edge rates for which inductance impacts delay tr 2 LC <l < 2L RC (4.50) As the example illustrated, inductance will only be important to the delay of lowresistance signals such as wide clock lines or the power supply. As edge rates become faster, inductance will become relevant to a larger number of on-chip signals. Inductive crosstalk is also important for wide busses far away from their current return paths. In power distribution networks, inductance means that if one portion of the chip requires a rapidly increasing amount of current, that charge must be delivered from nearby decoupling capacitors or supply pins; portions of the chip further away are unaware of the INTERCONNECT 123 Weste04r4.fm Page 124 Monday, January 5, 2004 1:59 AM 124 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION changing current needs until a speed-of-light flight time has elapsed and hence will not supply current immediately. Adding inductance to the power grid simulation generally reveals greater supply noise than would otherwise be predicted. Power supplies will be discussed further in Section 12.3. In wide, thick, upper-level metal lines, resistance and RC delay may be small. This pushes the horizontal line in Figure 4.45 upward, increasing the range of edge rates for which inductance matters. This is especially common for clock signals. Inductance tends to increase the propagation delay and sharpen the edge rate. To see the effects of inductance, consider a 5 mm metal6 clock line above a metal5 ground plane driving a 2pF clock load. If its width is 4.8 µm (six times minimum), it has resistance of 4 Ω/mm, capacitance of 0.4 pF/mm, and inductance of 0.12 nH/mm. Figure 4.46 presents models of the clock line as a 5-stage π-model without (a) and with (b) inductance. Figure 4.46(c) shows the response of each model to an ideal voltage source with 80 ps rise time. The model including inductance shows a greater delay until the clock begins to rise because of the speed of light flight time. It also overshoots. However, the rising edge is sharper and the rise time is shorter. In some circumstances when the driver impedance is matched to the characteristic impedance of the wire, the sharper rising edge can actually result in a shorter propagation delay measured at the 50% point. 4 4 0.4 pF 0.4 pF 4 0 .2 pF 4 4 0.4 pF 0 .4 pF rc 0 .2 pF 2 pF (a ) 4 0.12 nH 4 0 .2 pF 0.12 nH 4 0.4 pF 0.12 nH 4 0 .4 pF 0.12 nH 4 0 .4 pF 0 .4 pF (b) 2.0 1.5 Vin RC 1.0 V RLC 0.5 0 (c) t (ps) 0 200 400 0.12 nH 600 FIG 4.46 Wide clock line modeled with and without inductance 0 .2 pF lrc 2 pF Weste04r4.fm Page 125 Monday, January 5, 2004 1:24 AM 4.5 To reduce the inductance and the impact of skin effect when no ground plane is available, it is good practice to split wide wires into thinner sections interdigitated with power and ground lines to serve as return paths. For example, Figure 4.47 shows how a 16 µm wide clock line can be split into four 4 µm lines to reduce the inductance. 16 µm C LK (a) 4 µm C LK (b) G ND C LK VD D C LK G ND C LK VD D G ND FIG 4.47 Wide clock line interdigitated with power and ground lines to reduce inductance A bus made of closely spaced wires far above a ground plane is especially susceptible to inductive crosstalk. If all but one wire in the bus rises, each loop induces a magnetic field. These magnetic fields all pass through the loop formed by the nonswitching wire, in turn inducing a current in the victim wire. The noise from each aggressor sums on to the victim in much the same way that multiple primary turns in a transformer couple onto a single secondary turn. Computing the inductive crosstalk requires extracting a mutual inductance matrix for the bus and simulating the system. As this is not yet practical for large chips, designers instead either follow design rules that keep the inductive effects small or ignore inductance and hope for the best. The design rules may be of the form that one power or ground wire must be inserted between every N signal lines on each layer. N is called the signal:return (SR) ratio [Morton99]. N = 4 eliminates most inductive effects on noise and delay in a 180 nm process. N = 2 means each signal is shielded on one side, also eliminating half the capacitive crosstalk. However, low SR ratios are expensive in terms of metal resources. In summary, on-chip inductance is difficult to extract. Mutual inductive coupling may occur over a long range, so inductive coupling is difficult to simulate even if accurate values are extracted. Instead, design rules are usually constructed so that inductive effects may be neglected for most structures. A regular power and ground grid or plane with no gaps is essential. Power and ground lines should be interdigitated in wide signal lines such as clocks and between about every four bits in large high-speed busses. Inductance should be incorporated into simulations of the power and clock networks and into the noise and delay calculations for busses with large SR ratios in high-speed designs. INTERCONNECT 125 Weste04r4.fm Page 126 Monday, January 5, 2004 1:24 AM 126 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION 4.5.6 Temperature Dependence Interconnect capacitance is independent of temperature, but the resistance varies strongly. The temperature coefficients of copper and aluminum are about 0.4%/°C over the normal operating range of circuits; that is, a 100° C increase in temperature leads to 40% higher resistance. At liquid nitrogen temperature, the bulk resistivity of copper drops to 0.22 µΩcm, an eight-fold improvement. This suggests great advantages for RC-dominated paths in cooled systems. 4.5.7 An Aside on Effective Resistance and Elmore Delay In this chapter we have played fast and loose with the relationship between resistance and delay. In practice, the methods work well for hand estimations. When a highly accurate result is desired, you can simulate the circuit with actual device models and wire parameters rather than simplified RC circuits. Nevertheless, rigor compels us to revisit this relationship. According to the Elmore delay model, a gate with effective resistance R and capacitance C has a propagation delay of RC. A wire with distributed resistance R and capacitance C treated as a single π -segment has propagation delay RC /2. Reviewing the properties of RC circuits, we recall that the lumped RC circuit in Figure 4.48(a) has a unit step response of V out (t ) = 1 − e R ′C −t (4.51) The propagation delay of this circuit is obtained by solving for tpd when Vout(tpd) = 1/2: t pd = R ′C ln 2 = 0.69R ′C (4.52) The distributed RC circuit in Figure 4.48(b) has no closed form time domain response. Because the capacitance is distributed along the circuit rather than all being at the end, you would expect the capacitance to be charged on average through about half the R' Vin (t) C Vout (t) V (t) out 1 (a ) 0.5 R' Vin (t) (b) Distributed t pd C Vout (t) 0 Lumpe d 0.5 0 1 (c) FIG 4.48 Lumped and distributed RC circuit response 1.5 2 2.5 3 t R'C Weste04r4.fm Page 127 Monday, January 5, 2004 1:24 AM 4.5 resistance and that the propagation delay should thus be about half as great. A numerical analysis finds that the propagation delay is 0.38R′C. To reconcile the Elmore model with the true results for a logic gate, recall that logic gates have complex nonlinear I-V characteristics and are approximated as having an effective resistance. If we characterize that effective resistance as R = R′ln2, the propagation delay really becomes the product of the effective resistance and the capacitance: tpd = RC. In Section 5.4.5, we will calculate this effective resistance by simulating the delay of a gate driving a capacitive load and measuring the propagation delay. For distributed circuits, observe that 0.38R ′C ≈ 1 R ′C ln 2 = 1 RC . 2 2 Therefore, the Elmore delay model describes distributed delay well if we use an effective wire resistance equal to 69% of that computed with EQ (4.34) is used. This is somewhat inconvenient. The effective resistance is further complicated by the effect of nonzero rise time on propagation delay. Figure 4.49 shows that the propagation delay depends on the risetime of the input and approaches RC for lumped systems and RC/2 for distributed systems when the input is a slow ramp. This suggests that when the input is slow, the effective resistance for delay calculations in a distributed RC circuit is equal to the true resistance. Finally, we note that for many analyses such as repeater insertion calculations in Section 4.6.4, the results are only weakly sensitive to wire resistance, so using the true wire resistance does not introduce great error. V(t) Input Distributed 1 Lumped 0.5 (a ) 0 0 1 t pd 1 RC 2 3 t 4 R'C 3 4 Lumped ln 2 Distributed 0.5 0.38 (b) 0 0 1 2 FIG 4.49 Effect of rise time on lumped and distributed RC circuit delays t rise RC INTERCONNECT 127 Weste04r4.fm Page 128 Monday, January 5, 2004 1:24 AM 128 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION In summary, it is a reasonable practice to estimate propagation delay of gates using the Elmore delay model as RC where R is the effective resistance of the gate. Similarly, you can estimate the flight time along a wire as RC/2 where R is the true resistance of the wire. When more accurate results are needed, it is important to use good transistor models and appropriate input slopes in simulation. On a separate topic, Section 4.2.1 defined the Elmore delay of an RC ladder. In general, a wire branching to many destinations is usually modeled as an RC tree instead. The Elmore delay to node i of an RC tree is N TD i = ∑ RkiC k (4.53) k =1 where N is the number of nodes in the tree, Ck is the capacitance on node I, and Rki is the resistance between the input and node k in common with the path between the input and node i. This simplifies to the simple product of resistance and capacitance for each node in an RC ladder. In trees, the capacitance on branches away from the path to the output is conservatively lumped as if it were at the branch point on the path. Example Figure 4.50 models a gate driving wires to two destinations. The gate is represented as a voltage source with effective resistance R1. The two receivers are located at nodes 3 and 4. The wire to node 3 is long enough that it is represented with a pair of π-segments, while the wire to node 4 is represented with a single segment. Find the Elmore delay from input x to each receiver. Solution: The Elmore delays are: TD 3 = R1C 1 + (R1 + R 2 )C 2 + (R1 + R 2 + R3 )C 3 + R1C 4 (4.54) TD 4 = R1C 1 + R1C 2 + R1C 3 + (R1 + R 4 )C 4 R4 Medium Wire x (a ) C4 Node 4 Long Wire R1 R2 R3 C1 C2 Node 3 (b) FIG 4.50 Interconnect modeling with RC tree C3 Weste04r4.fm Page 129 Monday, January 5, 2004 1:24 AM 4.6 WIRE ENGINEERING The Elmore delay can be viewed in terms of the first moment of the impulse response of the circuit. CAD tools can obtain greater accuracy by approximating delay based on higher moments using a technique called moment matching. Asymptotic Waveform Evaluation (AWE) uses moment matching to estimate interconnect delay with better accuracy than the Elmore delay model and faster run times than a full circuit simulation [Celik02]. 4.6 Wire Engineering As gate delays continue to improve while long wire delays remain constant or even get slower, wire engineering has become a major part of integrated circuit design. It is necessary to develop a floorplan early in the design cycle, identify the long wires, and plan for them. While floorplanning in such a way that critical communicating units are close to one another has the greatest impact on performance, it is inevitable that long wires will still exist. The designer has a number of techniques to engineer wires for delay and coupling noise. The width, spacing, and layer usage are all under the designer’s control. Shielding can be used to further reduce coupling on critical nets. Repeaters inserted along long wires reduce the delay from a quadratic to a linear function of length. Wire capacitance and resistance complicate the use of Logical Effort in selecting gate sizes. 4.6.1 Width and Spacing The designer selects the wire width, spacing, and layer usage to achieve acceptable delay and noise. By default, minimum pitch wires are preferred for noncritical interconnections for best density. Widening a wire proportionally reduces resistance but increases the capacitance of its top and bottom plates. Table 4.8 showed that this leads to less than a proportional increase in capacitance, so the RC delay product improves, especially for narrow wires. Widening wires also increases the fraction of capacitance of the top and bottom plates, which somewhat reduces coupling noise from adjacent wires. Increasing spacing between wires reduces capacitance to the adjacent wires and leaves resistance unchanged. This improves the RC delay to some extent and significantly reduces coupling noise. Figure 4.41(a) shows the RC delay of a 10 mm metal2 wire in the process from Section 4.5 sandwiched between metal1 and metal3 planes as a function of wire pitch (w + s) for different wire spacings. Figure 4.51(b) shows the coupling capacitance to the two adjacent neighbors as a fraction of the total capacitance. The data shows that for tight pitches, it is better to increase width than spacing to improve delay. However, it is better to increase spacing than width to improve coupling. The best tradeoff clearly depends on the situation. 4.6.2 Layer Selection Early (1970s) MOS processes offered only a single layer of metal. Polysilicon or diffusion jumpers were required when two signals crossed. Modern processes have six or more metal layers. The lower layers are thin and optimized for a tight routing pitch. Middle layers are 129 Weste04r4.fm Page 130 Monday, January 5, 2004 1:24 AM 130 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION 0.8 0.7 Coupling: 2Cadj / (2Cadj+Cgnd) 2.0 1.8 Delay (ns): RC/2 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.6 0.4 0.3 0.2 0.1 0 0 (a) Wire Spacing (nm) 320 480 640 0.5 0 500 1000 Pitch (nm) 1500 2000 (b) 0 500 1000 1500 2000 Pitch (nm) FIG 4.51 Delay and coupling of 10 mm metal2 wire for various wire pitches and spacings often slightly thicker for lower resistance and better current-handling capability. Upper layers may be even thicker to provide a low-resistance power grid and fast global interconnect. A large number of layers are important to be able to route complex chips and supply power, ground, and the clock. Additional layers are also valuable because they allow a relaxed wire pitch, reducing RC delay and coupling problems. Wiring tracks are a precious resource and are often allocated in the floorplan; the wise designer maintains a reserve of wiring tracks for unanticipated changes late in the design process. A sample allocation of wiring tracks for a six-layer metal process is given in Table 4.11. Critical signals within a unit can be assigned upper-level metal tracks to reduce delay. Table 4.11 Sample metal layer usage in 6-level process Layer Purpose Metal 1 Metal 2/3 Metal 4/5 Metal 6 Interconnect within cells Interconnect between cells within units Interconnect between units, critical signals I/O pads, clock, power, ground The power grid is usually distributed over multiple layers. Most of the current-handling capability is provided in the upper two layers with lowest resistance. However, the grid must extend down to metal1 or metal2 to provide easy connection to cells. There is debate over the best use of the growing number of wiring layers to solve inductive and capacitive noise problems. One approach is to dedicate power and ground Weste04r4.fm Page 131 Monday, January 5, 2004 1:24 AM 4.6 WIRE ENGINEERING planes, much like in a printed circuit board. For example, the Alpha 21264 [Gronowski98] used thick metal3 and metal6 planes for ground and power. This was necessary because the I/O pins were all along the periphery of the chip and the planes were needed to provide a low-resistance path for power to the center; better packaging with power and ground bumps across the die reduces the pressure to use planes. Another approach is to shield noise-sensitive nets with power or ground lines; this will be discussed in Section 4.6.3. A third approach is to use differential signaling. 4.6.3 Shielding As discussed in Section 4.5.4, coupling from adjacent lines impacts both the delay and signal integrity of wires. The coupling can be avoided if the adjacent lines do not switch. It is common practice to shield critical signals with power or ground wires on one or both sides to eliminate coupling. This is costly in area but may be less costly than increasing spacing to the point that the coupling is negligible. For example, clock wires are usually shielded so that switching neighbors do not affect the delay of the clock wire and introduce clock skew. Sensitive analog wires passing near digital signals should also be shielded. An alternative to shielding is to interdigitate wires that are guaranteed to switch at different times. For example, if bus A switches on the rising edge of the clock and bus B switches on the falling edge of the clock, by interleaving the bits of the two busses you can guarantee that both neighbors are constant during a switching event. This avoids the delay impact of coupling; however, you must still ensure that coupling noise does not exceed noise budgets. Figure 4.52 shows wires shielded (a) on one side, (b) on both sides, and (c) interdigitated. Very sensitive signals such as clocks or analog voltages can be shielded above and below as well. 4.6.4 Repeaters Both resistance and capacitance increase with wire length l, so the RC delay of a wire increases with l 2, as shown in Figure 4.53(a). The delay may be reduced by splitting the wire into N segments and inserting an inverter or buffer called a repeater to actively drive the wire [Glasser85], as shown in Figure 4.53(b). The new wire involves N segments with RC flight time of (l/N)2, for a total delay of l2/N. If the number of segments is proportional to the length, the overall delay increases only linearly with l. vdd a0 a1 gnd a2 a3 vdd (a ) FIG 4.52 Wire shielding topologies vdd a0 gnd a1 vdd a2 gnd (b) a0 b0 a1 b1 (c) a2 b2 131 Weste04r4.fm Page 132 Monday, January 5, 2004 1:24 AM 132 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Wire Length: I (a ) Driver Receiver N Segments Segment l/ N (b) Driver l/ N Repeater l/ N Repeater Repeater Receiver FIG 4.53 Wire with and without repeaters Using inverters as repeaters gives best performance. Each repeater adds some delay. If the distance is too great between repeaters, the delay will be dominated by the long wires. If the distance is too small, the delay will be dominated by the large number of inverters. As usual, the best distance between repeaters is a compromise between these extremes. Suppose a unit inverter has resistance R and capacitance C ′ (C ′ ≈ 3C because the inverter is composed of a unit nMOS and double-width pMOS8) and a wire has resistance Rw and capacitance Cw per unit length. Consider inserting repeaters of W times unit size. You can show (see Exercise 33) that under the Elmore delay model, neglecting diffusion parasitics, the best length of wire between repeaters is l = N 2RC ′ RwC w (4.55) The delay per unit length of a properly repeated wire is t pd l ( = 2+ 2 ) RC ′RwC w (4.56) To achieve this delay, the inverters should use an nMOS transistor width of W= RC w RwC ′ (4.57) Unfortunately, inverting repeaters complicate design because you must either ensure an even number of repeaters on each wire or adapt the receiving logic to accept an inverted 8 It is not necessary to use equal rise and fall resistances and Section 6.2.1.6 shows that the delay and area can be slightly improved using smaller pMOS transistors. Weste04r4.fm Page 133 Monday, January 5, 2004 1:24 AM 4.6 WIRE ENGINEERING input. Some designers use inverter pairs (buffers) rather than single inverters to avoid the polarity problem. The pairs contribute more delay. However, the first inverter size W1 may be smaller, presenting less load on the wire driving it. The second inverter may be larger, driving the next wire more strongly. You can show that the best size of the second inverter is W2 = kW1, where k = 2.06 if diffusion parasitics are negligible. The distance between repeaters increases to (see Exercise 34) 1 2RC ′ k + k RwC w l = N (4.58) The delay per unit length becomes t pd l = 3.65 RC ′RwC w (4.59) using transistor widths of W1 = W k , W2 = W k (4.60) This typically means that wires driven with noninverting repeaters are only about 7% slower per unit length than those using inverting repeaters. Only about two-thirds as many repeaters are required, simplifying floorplanning. Total repeater area and power increases slightly. When diffusion parasitics are considered, each repeater becomes slightly more expensive. Therefore, it is better to use fewer repeaters spaced at greater distances, as you would expect from Logical Effort. The best transistor sizes are unchanged. There are no closedform results, but the problem can be solved numerically or thorough simulation. The overall delay is a weak function of the distance between repeaters, so it is reasonable to increase this distance to reduce the difficulty of finding places in the floorplan for repeaters while only slightly increasing delay. Repeaters impose directionality on a wire. Bidirectional busses and distributed tristate busses cannot use simple repeaters and hence are slower; this favors point-to-point unidirectional communications. Repeaters also draw unusually large crowbar currents because their inputs come from RC lines with slow edge rates, turning both transistors partially ON for a long time. An alternative to repeaters are boosters [Nalamalpu02], which are placed in parallel rather than in series with a wire as shown in Figure 4.54(a). The booster senses when a wire is switching and aids the transition. It relies on the principles of hysterisis and positive feedback to allow bidirectional operation at the expense of reduced noise margins. The 133 Weste04r4.fm Page 134 Monday, January 5, 2004 1:24 AM 134 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Example Determine the best distance between repeaters for a minimum pitch metal2 line in a 180 nm process. Assume the metal usage on other layers is dense. The transistor resistance is 3KΩ • µm and the gate capacitance is C = 1.7 fF/µm. How far should the repeaters be spaced? How wide should the repeater transistors be? What is the signal velocity on the wire? How do your results change if width and spacing are each increased by 50%? How do they change for a minimum-pitch metal5 line? Solution: The capacitance of minimum-pitch metal2 lines with planes above and below is Cw = 0.21 fF/µm, according to Table 4.8. The sheet resistance from Table 4.7 is 0.05 Ω/square, or Rw = Ω 0.05 square µm 0.32 square = 0.16 µΩ m (4.61) C ′ = 3C = 5.1 fF/µm. According to EQ (4.55), the distance between repeaters should be ( fF 2(3000Ω • µm ) 5.1 µm (0.16 )( Ω µm fF 0.21 µm ) ) = 950 µm (4.62) According to EQ (4.57), the nMOS transistor width should be (3000Ω • µm)(0.21 µfF ) m (0.16 )(5.1 ) Ω µm fF µm = 28 µm (4.63) and the pMOS should be roughly twice that. The repeated signal velocity (ignoring diffusion capacitance) is (2 + 2 ) ps (3000Ω • µm)(5.1 µfF )(0.16 µΩ )(0.21 µfF ) = 71 mm m m m (4.64) Weste04r4.fm Page 135 Monday, January 5, 2004 1:24 AM 4.6 WIRE ENGINEERING In comparison, the speed of light in a medium with a relative permittivity of 3.55 is 3 • 108 m s 3.55 ps = 0.155 mm = 6.3 mm ps (4.65) Even with repeaters, the wire is an order of magnitude slower than the speed of light because it is so resistive. If the width and spacing increase by 50%, the capacitance is 0.18 fF/µm and the resistance is Rw = Ω 0.05 square µm 1.5 • 0.32 square = 0.10 µΩ m (4.66) As the wire is faster, we would expect that the repeaters could be placed further apart but that larger transistors would be needed to drive the longer segments. Reevaluating the same equations gives a distance of 1300 µm between repeaters, transistor widths of 33 µm, and wire delay of 52 ps/mm. A metal5 line is even faster with a capacitance of 0.24 fF/µm and resistance of Rw = Ω 0.02 square µm 0.8 square = 0.025 µΩ m (4.67) The best repeater spacing is 2260 µm with 75 µm transistors, giving a delay of 30 ps/mm. It is clear that the fat, wide upper level metal lines are a precious routing resource. gate in the center is an inverting Muller C-element, defined in Figure 4.54(b), which provides hysteresis by only toggling when both inputs are the same. Skewed inverters with high and low switching points sense the beginning of a transition. Figure 4.54(c) shows simulations of the booster in action on a 6mm long metal2 wire. The wire is initially ‘0,’ so l and h are initially ‘1’ and c is initially ‘0.’ In this state, the booster is OFF and the weak 135 Weste04r4.fm Page 136 Monday, January 5, 2004 2:50 AM 136 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION keeper holds the wire low. The input rises and the middle of the wire slowly follows. The keeper opposes the transition but is weak enough to have little effect. When the middle reaches the switching point of the LO-skew inverter, node l falls. Now the booster fires through both pMOS transistors, serving as positive feedback to strongly pull the middle of the wire high. This is visible as a kink in the mid waveform. The keeper also turns OFF. As the middle reaches the switching point of the HI-skew inverter, node h falls. c rises and turns OFF the booster because the transition is nearly complete. The keeper turns ON again, holding the middle of the wire high. The waveforms without boosters are also shown; they are slowly rising exponentials. Nalamalpu finds boosters are 20% faster and are inserted at three times the spacing of repeaters. [Dobbalaere95] presents a similar technique using self-timed circuits. Wire Wire Mid In Out l L A C H A B c C h (a) (b) Weak Keeper Strong Booster 2.0 In Mid Out l h c Mid (No Booster) Out (No Booster) 1.5 1.0 V 0.5 (c) 0 t (ns) 0 FIG 4.54 Booster 1 2 C= B C Weste04r4.fm Page 137 Monday, January 5, 2004 1:24 AM 4.6 4.6.5 WIRE ENGINEERING Implications for Logical Effort Interconnect complicates the application of Logical Effort because the wires have a fixed capacitance. The branching effort at a wire with capacitance Cwire driving a gate load of Cgate is (Cgate + Cwire) / Cgate. This branching effort is not constant; it depends on the size of the gate being driven. The simple rule that circuits are fastest when all stages bear equal effort is no longer true when wire capacitance is introduced. If the wire is very short or very long, there are simple approximations possible, but when the wire and gate loads are comparable, there is no simple method to determine the best stage effort. Every circuit has some interconnect, but when the interconnect is short (Cwire << Cgate), it can be ignored. Alternatively, you can compute the average ratio of wire capacitance to parasitic diffusion capacitance and add this as extra parasitic capacitance when determining parasitic delay. For connections between nearby gates, this generally leads to a best stage effort ρ slightly greater than 4. Conversely, when the interconnect is long (Cwire >> Cgate), the gate at the end can be ignored. The path can now be partitioned into two parts. The first part drives the wire while the second receives its input from the wire. The first part is designed to drive the load capacitance of the wire; the extra load of the receiver is negligible. A long wire is often driven by an inverter with a stage effort of 8 to 12 rather than 4 because low stage efforts require large, power-hungry drivers and have little performance advantage when the delay is dominated by the wire RC. The size of the receiver is chosen by practical considerations: Larger receivers may be faster, but they also cost area and power. The most difficult problems occur when Cwire ≈ Cgate. These medium-length wires introduce branching efforts that are a strong function of the size of the gates they drive. Writing a delay equation as a function of the gate sizes along the path and the wire capacitance results in an expression that can be differentiated with respect to gate sizes to compute the best sizes. For the purpose of hand estimation, this is usually too much work. A reasonable alternative is to preserve the stage effort of about four. The initial branching effort of the wire is unknown, so iteration is usually necessary. 4.6.6 Crosstalk Control Recall from EQ (4.40) that the capacitive crosstalk is proportional to the ratio of coupling capacitance to total capacitance. For modern wires with an aspect ratio (t/s) of 2 or greater, the coupling capacitance can account for two-thirds or more of the total capacitance and crosstalk can create large amounts of noise and huge data-dependent delay variations. Shielding or increasing the spacing or width of wires alleviates crosstalk at the expense of greater area. Three alternative techniques to control crosstalk are staggered repeaters, charge compensation, and twisted differential signaling [Ho03]. Each technique seeks to cause equal amounts of positive and negative crosstalk on the victim, effectively producing zero net crosstalk. Figure 4.56(a) shows two wires with staggered repeaters. Each segment of the victim sees half of a rising aggressor segment and half of a falling aggressor segment. Although the cancellation is not perfect because of delays along the segments, staggering is a very 137 Weste04r4.fm Page 138 Monday, January 5, 2004 1:24 AM 138 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Example The path in Figure 4.55 contains a medium-length wire modeled as a lumped capacitance. Write an equation for path delay in terms of x and y. How large should the x and y inverters be for shortest path delay? What is the stage effort of each stage? Solution: From the Logical Effort delay model, we find the path delay is d= x y + 50 100 + + 10 x y (4.68) Differentiating with respect to each size and setting the results to 0 allows us to solve EQ (4.40) for x = 31 fF and y = 56 fF. The stage efforts are (31/10) = 3.1, (56 + 50)/33 = 3.1, and (100/56) = 1.8. Notice that the first two stage efforts are equal as usual, but 10 fF x y the third stage effort is lower. As x already 50 fF 100 fF drives a large wire capacitance, y may be rather large (and will bear a small stage effort) before the incremental increase in delay of x driving y FIG 4.55 Path with mediumequals the incremental decreases in delay of y length wire driving the output. 1 y + 50 − = 0 ⇒ x 2 = 10 y + 500 10 x2 1 100 − 2 = 0 ⇒ y 2 = 100x x y (4.69) effective approach. Figure 4.56(b) shows charge compensation in which an inverter and transistor are added between the aggressor and victim. The transistor is connected to behave as a capacitor. When the aggressor rises and couples the victim upward, the inverter falls and couples the victim downward. By choosing an appropriately sized compensation transistor, most of the noise can be cancelled at the expense of the extra circuitry. Figure 4.56(c) shows twisted differential signaling in which each signal is routed differentially. The signals are swapped or twisted such that the victim and its complement each see equal coupling from the aggressor and its complement. This approach is expensive in wiring resources, but it very effectively eliminates crosstalk. It is widely used in memory designs, as explored in Section 11.2.3. Weste04r4.fm Page 139 Monday, January 5, 2004 1:24 AM 4.6 WIRE ENGINEERING Victim Coupled Noise Cancels Aggressor (a ) Victim Coupled Noise Cancels Aggressor (b) v v a (c) a FIG 4.56 Crosstalk control schemes 4.6.7 Low-swing Signaling Driving long wires is slow because of the RC delay, and expensive in power because of the large capacitance to switch. Low-swing signaling improves performance by sensing when a wire has swung through some small Vswing rather than waiting for a full swing. If the driver is turned off after the output has swung sufficiently, the power can be reduced as well. However, the improvements come at the expense of more complicated driver and receiver circuits. Low-swing signaling may also require a twisted differential pair of wires to eliminate common-mode noise that could corrupt the small signal. The power consumption for low-swing signaling depends on both the driver voltage Vdrive and the actual voltage swing Vswing. Each time the wire is charged and discharged, it consumes Q = CVswing. If the effective switching frequency of the wire is αf, the average current is T I avg = 1 idrive (t ) = αfCV swing T∫ 0 (4.70) 139 Weste04r4.fm Page 140 Monday, January 5, 2004 1:24 AM 140 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Hence, the dynamic dissipation is Pdynamic = I avgV drive = αfCV swingV drive (4.71) In contrast, a rail-to-rail driver uses Vdrive = Vswing = VDD and thus consumes power proportional to VDD2 (EQ (4.30)). Vswing must be less than or equal to Vdrive. By making Vswing less than Vdrive, we speed up the wire because we do not need to wait for a full swing. By making both voltages significantly less than VDD, we can reduce the power by an order of magnitude. Low-swing signaling involves numerous challenges. A low Vdrive must be provided to the chip and distributed to low-swing drivers. The signal should be transmitted on differential pairs of wires that are twisted to cancel coupling from neighbors and equalized to prevent interference from the previous data transmitted. The driver must turn on long enough to produce Vswing at the far end of the line, then turned off to prevent unnecessary power dissipation. This generally leads to a somewhat larger swing at the near end of the line. The receiver must be clocked at the appropriate time to amplify the differential signal. Distributing a self-timed clock from driver to receiver is difficult because the distances are long so the time to transmit a full-swing clock exceeds the time for the data to complete its small swing. [Ho03] describes a synchronous low-swing signaling technique using the system clock for both driver and receiver. During the first half of the cycle, the driver is OFF (high impedance) and the differential wires are equalized to the same voltage. During the second half of the cycle, the drivers turn ON. At the end of the cycle, the receiver senses the differential voltage and amplifies it to full-swing levels. Figure 4.57(a) shows the overall system architecture. Figure 4.57(b) shows the driver for one of the wires. The gates use ordinary VDD while the drive transistors use Vdrive. Because Vdrive << VDD, nMOS transistors are used for both the pullup and pulldown to deliver low effective resistance in their linear regime. A second driver using the complementary input drives the complementary wire. Figure 4.57(c) shows the differential wires with twisting and equalizing. Figure 4.57(d) shows the clocked sense amplifier based on the SA-F/F that will be described further in Section 7.3.8. The sense amplifier uses pMOS input transistors because the small-swing inputs are close to GND and below the threshold of nMOS transistors. The design in [Ho03] used a 1.8 V VDD, 0.65 V Vdrive, and 0.1 V minimum Vswing in a 180 nm process. The clock frequency was 1 GHz, or 10 FO4 inverter delays, to drive a 10 mm wire. Note that the clock period must be long enough to transmit an adequate voltage swing. If the clock period increases, the circuit will actually dissipate more power because the voltage swing will increase to a maximum of Vdrive. Weste04r4.fm Page 141 Monday, January 5, 2004 1:24 AM 4.7 clk clk Low Sw ng i Driver a clk Twisted, Equalized Differential Wires Sense Amplifier y y (a) Vdrive Q SR Latch R clk a S VDD d clk f (b) d f f clk clk clk (c) d clk f (d) FIG 4.57 Low-swing signaling system 4.7 Design Margin So far, when considering the various aspects of determining a circuit’s behavior, we have only alluded to the variations that might occur in this behavior given different operating conditions. In general, there are three different sources of variation—two environmental and one manufacturing. These are: Supply voltage Operating temperature Process variation You must aim to design a circuit that will reliably operate over all extremes of these three variables. Failure to do so invites circuit failure, potentially catastrophic system failure, and a rapid decline in reliability (not to mention a loss of customers). Variations can be modeled with uniform or normal (Gaussian) statistical distributions. These distributions are shown in Figure 4.58. Uniform distributions are specified with a half-range. For good results, accept variations over the entire half-range. For example, VDD may be specified at 1.2 V +/– 10%. This is a uniform distribution with a 120 mV halfrange and all parts should work at any voltage in the range. Normal distributions are specified with a standard deviation σ. Processing variations are usually modeled with normal distributions. Retaining parts with a 3σ distribution will result in 0.26% of parts being rejected. A 2σ retention results in 4.56% of parts being rejected, while 1σ results in a 31.74% rejection rate. Obviously, rejecting parts outside 1σ of nominal would waste a large number of parts. A 3σ or 2σ limit is conventional and a manufacturer with a com- DESIGN MARGIN 141 Weste04r4.fm Page 142 Monday, January 5, 2004 1:24 AM 142 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION 1 1 All parts lie within the half-range Accepting parts within 1s would exclude 31.7% Accepting parts within 2s would exclude 4.6% 1/e Accepting parts within 3s would exclude 0.3% -3 -2 -1 0 1 Normal Distribution 2 3 -1 0 1 Uniform Distribution FIG 4.58 Uniform and normal distributions mercially viable CMOS process should be able to supply a set of device parameters that are guaranteed to yield at this rate. However, as the variations are getting more significant, designers are moving toward statistical rather than worst-case design [Borkar03]. 4.7.1 Supply Voltage Systems are designed to operate at a nominal supply voltage, but this voltage may vary for many reasons including tolerances of the voltage regulator, IR drops along supply rails, and di/dt noise. The system designer may tradeoff power supply noise against resources devoted to power supply regulation and distribution; typically the supply is specified at +/– 10% around nominal at each logic gate. In other words, the variation has a uniform distribution with a half-range of 10% of VDD. Speed is roughly proportional to VDD, so to first order this leads to +/– 10% delay variations (check for your process and voltage when this is critical). Power supply variations also appear in noise budgets. 4.7.2 Temperature Section 2.2.3.7 showed that as temperature increases, drain current decreases. The junction temperature of a transistor is the sum of the ambient temperature and the temperature rise caused by power dissipation in the package. This rise is determined by the power consumption and the package thermal resistance, as discussed in Section 12.2.4. Table 4.12 lists the ambient temperature ranges for parts specified to commercial, industrial, and military standards. Parts must operate at the bottom end of the ambient range unless they are allowed time to warm up before use. The junction temperature may significantly exceed the maximum ambient temperature. Commonly commercial parts are verified to operate with junction temperatures up to 110° to 125° C. Weste04r4.fm Page 143 Monday, January 5, 2004 1:24 AM 4.7 Table 4.12 Ambient temperature ranges Standard 4.7.3 Minimum Maximum Commercial Industrial Military 0° C –40° C –55° C 70° C 85° C 125° C Process Variation Devices and interconnect have variations in film thickness, lateral dimensions, and doping concentrations [Bernstein98]. These variations occur from one wafer to another, between dice on the same wafer, and across an individual die; variation is generally smaller across a die than between wafers. These effects are sometimes called inter-die and intra-die variations; intra-die variation is also called process tilt because certain parameters may slowly and systematically vary across a die. For example, if an ion implanter delivered a greater dose nearer the center of a wafer than near the periphery, the threshold voltages might tilt radially across the wafer. For devices, the most important variations are channel length L, oxide thickness tox, and threshold voltage Vt. Channel length variations are caused by photolithography proximity effects, deviations in the optics, and plasma etch dependencies. Oxide thickness is well controlled and generally is only significant between wafers; its effects on performance are often lumped into the channel length variation. Threshold voltages vary because of different doping concentrations and annealing effects, mobile charge in the gate oxide, and discrete dopant variations caused by the small number of dopant atoms in tiny transistors. For interconnect, the most important variations are line width and spacing, metal and dielectric thickness, and contact resistance. Line width and spacing, like channel length, depend on photolithography and etching proximity effects. Thickness may be influenced by polishing. Contact resistance depends on contact dimensions and the etch and clean steps. 4.7.4 Design Corners From the designer’s point of view, the collective effects of process and environmental variation can be lumped into their effect on transistors: typical (also called nominal), fast, or slow. In CMOS, there are two types of transistors with somewhat independent characteristics, so the speed of each can be characterized. Moreover, interconnect speed may vary independently of devices. When these processing variations are combined with the environmental variations, we define design or process corners. The term corner refers to an imaginary box that surrounds the guaranteed performance of the circuits, shown in Figure 4.59. The box is not square because some characteristics such as oxide thickness track between devices, making it impossible to simultaneously find a slow nMOS transistor with thick oxide and a fast pMOS transistor with thin oxide. DESIGN MARGIN 143 Weste04r4.fm Page 144 Monday, January 5, 2004 1:24 AM 144 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Slow pMOS Fast Table 4.13 lists a number of important design corners. The corners are specified with five letters describing the nMOS, pMOS, interconnect, power supply, and temperature, respectively. The letters FF SF are F, T, and S, for fast, typical, and slow. The environmental corners for a 1.8 V commercial process are shown in Table 4.14, illustrating that circuits are fastest at high voltage and low temperature. Circuits TT are most likely to fail at the corners of the design space, so nonstandard circuits should be simulated at all corners to ensure they operate correctly in all cases. Often integrated circuits are designed to meet a FS timing specification for typical processing. These parts may be binned; SS faster parts are rated for higher frequency and sold for more money, while slower parts are rated for lower frequency. In any event, the parts must still work in the slowest SSSSS environment. Other integrated Slow nMOS Fast circuits are designed to obtain high yield at a relatively low frequency; these parts are simulated for timing in the slow process corner. The FIG 4.59 Design corners fast corner FFFFF has maximum DC power dissipation because threshold voltages are lowest. Other corners are used to check for races and ratio problems where the relative strengths and speeds of different transistors or interconnect are important. The FFFFS corner is important for noise because the edge rates are fast, causing more coupling; the threshold voltages are low; and the leakage is high [Shepard99]. Often the corners are abbreviated to fewer letters. For example, two letters generally refer to nMOS and pMOS. Three refer to nMOS, pMOS, and environment. Four refer to nMOS, pMOS, voltage, and temperature. Table 4.13 Design corner checks Corner nMOS pMOS T T S S F F Purpose Wire T S F VDD S S F F S F S F S F F F F S T F S F F F S T F Temp S timing specifications (binned parts) S timing specifications (conservative) F DC power dissipation, race conditions, hold time constraints, pulse collapse, noise S subthreshold leakage noise, overall noise analysis S races of gates against wires F races of wires against gates F pseudo-nMOS & ratioed circuits noise margins, memory read/write, race of pMOS against nMOS F ratioed circuits, mem read/write, race of nMOS against pMOS Weste04r4.fm Page 145 Monday, January 5, 2004 1:24 AM 4.7 Table 4.14 Environmental corners Corner Voltage Temperature F T S 1.98 1.8 1.62 0° C 70° C 125° C It is important to know the design corner when interpreting delay specifications. For example, the datasheet in Figure 4.25 is specified at the 25° TTTT corner. The SS corner is 27% slower. The cells are derated at –71% per volt and 0.13%/°C, for additional penalties of 13% each in the low voltage and high temperature corners. These factors are multiplicative, giving SSSS delay of 1.62 times nominal. [Ho01] and [Chinnery02] find the FO4 inverter delay can be estimated from the effective channel length Leff as: Leff • (0.36 ps/nm) in TTTT corner Leff • (0.50 ps/nm) in TTSS corner Leff • (0.60 ps/nm) in SSSS corner Note that the effective channel length is aggressively scaled faster than the drawn channel length to improve performance. Typically Leff = 0.5-0.7 Ldrawn. For example, Intel’s 180 nm process was originally manufactured with Leff = 140 nm and eventually pushed to Leff = 100 nm. Our model predicts an FO4 inverter delay of about 70–50 ps in the TTSS corner where design usually takes place. In addition to working at the standard process corners, chips must function in a very high temperature, high voltage burn-in corner (e.g., 125–140° C externally, corresponding to an even higher internal temperature, and 1.3–1.7x nominal VDD [Vollersten99]), as described in Section 9. While it does not have to run at full speed, it must operate correctly so that all nodes can toggle. The burn-in corner has very high leakage and can dictate the size of keepers and weak feedback on domino gates and static latches. Processes with multiple threshold voltages and/or multiple oxide thicknesses can see each flavor of transistor independently varying as fast, typical, or slow. This can easily lead to more corners than anyone would care to simulate and raises challenges about identifying what corners must be checked for different types of circuits. 4.7.5 Matching There are many cases in which the designer cares how well two nominally identical transistors match. For example, in a sense amplifier, the minimum voltage that reliably can be sensed depends on the offset voltage of the amplifier, which in turn depends on mismatch between the input transistors. Differential pairs used in most analog circuits also are very sensitive to mismatch between input transistors. In a clock distribution network, we would like to distribute a clock to all points on a chip at the same time and mismatch leads to DESIGN MARGIN 145 Weste04r4.fm Page 146 Monday, January 5, 2004 1:24 AM 146 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION clock skew. It is clearly overkill to model one transistor as fast and an adjacent transistor as slow because nearby transistors experience similar processing; such conservative modeling would make design impossible. On the other hand, it is also clear that the two transistors are not identical. Designers should be able to obtain data about process variation from the manufacturer. Unfortunately, this data is often not available. Even if the manufacturer has characterized the process for variation, the data is usually a closely guarded trade secret. Moreover, manufacturers do not like to share the information for fear that it will become part of the process specification and prevent process modifications that lead to faster transistors at the expense of different variabilities. Nevertheless, device matching is a critical parameter for calculating clock skew and is an increasingly important specification for a process. Mismatches occur from both systematic variability and uncertainty [Naffziger02]. Systematic variability has a quantitative relationship to a source. For example, an ion implanter can systematically deliver a different dosage to different regions of a wafer. Similarly, polysilicon gates can systematically be etched narrower in regions of high polysilicon density than low density. Uncertainty occurs when the source is unknown, random, or too costly to model. Systematic variability can be modeled and nulled out; for example, in principle, you could examine a layout database and calculate the etching variations as a function of nearby layout, then simulate a circuit with suitably adjusted gate lengths. In practice, systematic variability often must be treated as uncertainty because of limited modeling resources. P rocess tilt describes the tendency for parameters to slowly vary from one corner of the die to another. Adjacent transistors tend to match better than widely separated ones. Variations in transistor threshold voltage and current have been found experimentally to scale with 1 / WL [Pelgrom89]. Device parameters depend on device size, orientation, and nearby polysilicon density. Therefore, for good matching it is best to build identical, relatively large transistors oriented in the same direction. For the same total gate area WL, long-channel transistors match better, so sense amplifiers can be built with longer than minimum devices [Lovett98]. When matching is especially critical, as in clock buffers, you can surround the transistor with a consistent pattern of polysilicon so all transistors see comparable nearby polysilicon densities. Threshold voltage variations are primarily caused by statistical fluctuations in the number of dopant atoms in the channel [Mizuno94]. As transistors shrink, they contain fewer dopant atoms and the statistical fluctuations become more severe. Matching problems can be characterized as systematic, random, drift, or jitter. Systematic mismatches include factors that can be modeled and simulated at design time, such as wires of different lengths. Random mismatches include most process variations (length, threshold, and interconnect) that are either truly random or too costly to model. These mismatches do not change with time, so they can be nulled out through feedback circuits that detect the variation and compensate. Drift mismatches, notably temperature variation, change slowly with time as compared to the operating frequency of the system. Drift can again be nulled by compensation circuits, but such circuits must sample repeatedly faster than the drift occurs rather than just once at manufacturing or startup. Jitter, often from voltage variations, is the most difficult cause of mismatch. It occurs at frequen- Weste04r4.fm Page 147 Monday, January 5, 2004 1:24 AM 4.7 cies comparable to or faster than the system clock and therefore may not be eliminated through feedback. As an example, [Harris01] presents process variation data used to model the clock distribution network of the Itanium 2 processor in the Intel 180 nm process [Yang98]. The primary sources of random mismatch are channel lengths and threshold voltages. The channel length has both a component slowly varying across the die and a random component with no apparent spatial correlation. The slowly varying component varies over a 12.5 nm half-range for transistors separated by 4 mm or more; nearby transistors see less variation and adjacent transistors see no variation. The random component is modeled with a standard deviation of 3.3 nm. The threshold voltage variation has a Gaussian distribution with an inverse area dependence. It is treated as 16.8 mV for small nMOS transistors, 14.6 mV for small pMOS transistors, 7.9 mV for large nMOS transistors, and 6.5 mV for large pMOS transistors. Large transistors are defined as those with width exceeding 12.5 µm. 4.7.6 Delay Tracking Another common design problem is how to build matched delays; for example, clockdelayed domino (see Section 7.5.4.2) needs to provide clocks to gates after their inputs have settled. The clocks must be matched to the gate delay; if they arrive late, the system functions slower, but if they arrive early, the system doesn’t work at all. Therefore, it is of great interest to the designer how well two delays can be matched. The best way to build matched delays is to actually provide replicas of the gates that are being matched. For example, in a static RAM (see Section 11.2.3), replica bitlines are used to determine when the sense amplifier should fire. Any relative variation in wire, diffusion, and gate capacitances happens to both circuits. In many situations, it is not practical to use replica gates; instead, a chain of inverters can be used. Unfortunately, even if there is no intra-die process variation, the inverter delay may not exactly track the delay it matches across design corners. For example, if the inverter chain were matching a wire delay in the typical corner, it would be faster than the wire in the FFSFF corner and slower than the wire in the SSFSS corner. This variation requires that the designer provide margin in the typical case so that even in the worst case, the matched delay does not arrive too early. How much margin is necessary? Figure 4.60 shows how gate delays, measured as a multiple of an FO4 inverter delay, vary with process, design corners, temperature, and voltage [Harris97]. The circuits studied include complementary CMOS NAND and NOR gates, domino AND and OR gates, and a 64-bit domino adder with significant wire RC delay. Figure 4.60(a) shows the gate delay of various circuits in different processes. The adder shows the greatest variation because of its wire-limited paths, but all the circuits track to within 20% across processes. This indicates that if a circuit delay is measured in FO4 inverter delays for one process, it will have a comparable delay in a different process. Figure 4.60(b and c) shows gate delay scaling with power supply voltage and temperature. Figure 4.60(d) shows what combination of design corner, voltage, and temperature gives the largest variation in delay normalized to an FO4 inverter in the same combination in the 0.6 µm process. Observe that the DESIGN MARGIN 147 Weste04r4.fm Page 148 Monday, January 5, 2004 1:24 AM 148 CHAPTER 4 (a ) (c) CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION (b) (d) FIG 4.60 Delay tracking variation is smallest for simple static CMOS gates that most closely resemble inverters and can reach 30% for some gates. These figures demonstrate that an inverter chain should have a nominal delay about 30% greater than the path it matches so that the inverter output always arrives later than the matched path across all combinations of voltage, temperature, and design corners. This is a hefty margin and discourages the casual use of matched delays. Weste04r4.fm Page 149 Monday, January 5, 2004 1:24 AM 4.8 4.8 Reliability Designing reliable CMOS chips involves understanding and addressing the potential failure modes [Greenhill02, Bernstein99]. This section addresses reliability problems (hard errors) that cause integrated circuits to fail permanently, including: Electromigration Self-heating Hot carriers Latchup O vervoltage failure This section also considers transient failures (soft errors) that cause the system to crash or lose data. Circuit pitfalls and common design errors are discussed in Section 6.3. 4.8.1 Reliability Terminology A number of acronyms are commonly used in describing reliability [Tobias95]. MTBF is the mean time between failures: (# devices • hours of operation) / # failures. FIT is the failures in time, the number of failures that would occur every thousand hours per million devices, or equivalently, 109 • (failure rate/hour). 1000 FIT is one failure in 106 hours = 114 years. This is good for a single chip. However, if a system contains 100 chips each rated at 1000 FIT and a customer purchases 10 systems, the failure rate is 100 • 1000 • 10 = 106 FIT, or one failure every 1000 hours (42 days). Reliability targets of less than 100 FIT are desirable. Most systems exhibit the bathtub curve shown in Figure 4.61. Soon after birth, systems with weak or marginal components tend to fail. This period is called infant mortality. Reliable systems then enter their useful operating life, in which the failure rate is low. Finally, the failure rate increases at the end of life as the system wears out. It is important to age systems past infant mortality before shipping the products. Aging is accelerated by stressing the part through burn-in at higher than normal voltage and temperature. Nodes must toggle during burn-in to stress all parts of the system. Therefore, the circuits must be functional even under burn-in conditions. Systems are also subjected to accelerated life testing during burn-in conditions to simulate the aging process and evaluate the time to wearout. The results are extrapolated to normal operating conditions to judge the actual useful operating life. This process is time-consuming and comes right at the end of the project. Part of any high-volume chip design will necessarily include designing a reliability assessment program that consists of burn-in boards deliberately stressing a number of chips over an extended period. Designers have tried to develop reliability simulators to predict lifetime [Hu92, Hsu92], but physical testing remains important. For high-volume parts, the source of failures is tracked and common points of failure can be redesigned and rolled into manufacturing. RELIABILITY 149 Weste04r4.fm Page 150 Monday, January 5, 2004 1:24 AM CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Failure Rate 150 Infant Mortality Useful Operating Life Wearout Time FIG 4.61 Reliability bathtub curve 4.8.2 Electromigration Electromigration causes wearout of metal interconnect through the formation of voids [Hu95]. High current densities lead to an “electron wind” that causes metal atoms to migrate over time. Remarkable videos taken under a scanning electron microscope show void formation and migration and wire failure [Meier99]. The problem is especially severe for aluminum wires; it is commonly alleviated with an Al-Cu or Al-Si alloy and is much less important for pure copper wires because of the different grain transport properties. The electromigration properties also depend on the grain structure of the metal film. Electromigration depends on the current density J = I/wt. It is more likely to occur for wires carrying a DC current where the electron wind blows in a constant direction than for those with bidirectional currents [Liew90]. Electromigration current limits are usually expressed as a maximum Jdc. The time to failure also is highly sensitive to operating temperature as given by Black’s Equation [Black69]: Ea MTTF ∝ e kT J dc n (4.72) where Ea is the activation energy that can be experimentally determined by stress testing at high temperatures and n is typically 2. The electromigration DC current limits vary with materials, processing, and desired MTTF and should be obtained from the fabrication vendor. In the absence of better information, a maximum Jdc of 1–2 mA/µm2 is a conservative limit for aluminum wires [Rzepka98], although 10 mA/µm2 or better may be achievable for copper wires [Young00]. Current density may be even more limited in contact cuts. Weste04r4.fm Page 151 Monday, January 5, 2004 2:06 AM 4.8 4.8.3 RELIABILITY 151 Self-heating While bidirectional wires are less prone to electromigration, their current density is still limited by self-heating. High currents dissipate power in the wire, which raises its ambient temperature. Hot wires exhibit greater resistance and delay. Electromigration is also highly sensitive to temperature, so self-heating may cause temperature-induced electromigration problems in the bidirectional wires. Brief pulses of high peak currents may even melt the interconnect. Self-heating is dependent on the RMS current density. This can be measured with a circuit simulator or calculated as Jdc T ∫ I (t ) dt VDD 2 I rms = 0 T A reasonable rule to control reliability problems with self-heating is to keep J r m s < 15 mA/ µ m 2 for bidirectional aluminum wires [Rzepka98] on a silicon substrate. Self-heating is especially significant for SOI processes because of the poor thermal conductivity of SiO2. In summary, electromigration from high DC current densities is primarily a problem in power and ground lines. Self-heating limits the RMS current density in bidirectional signal lines. However, do not overlook the significant unidirectional currents that flow through the wires contacting nMOS and pMOS transistors. For example, Figure 4.62 shows which lines in an inverter are limited by DC and RMS currents. Both problems can be addressed by widening the lines or reducing the transistor sizes (and hence the current). 4.8.4 Jdc (4.73) Jdc Jrms Jrms Jdc Jdc GND Jdc FIG 4.62 Current density limits in an inverter Hot Carriers As transistors switch, some high-energy (“hot”) carriers may be injected into the gate oxide and become trapped there. The damaged oxide changes the I-V characteristics of the device, reducing current in nMOS transistors and increasing current in pMOS transistors. Damage is maximized when the substrate current Isub is large, which typically occurs when nMOS transistors are in saturation while the input rises. Therefore, the problem is worst for inverters and NOR gates with slowly rising inputs and heavily loaded outputs [Sakurai86], and for high power supply voltages. Hot carriers cause circuit wearout as nMOS transistors become too slow. They can also cause failures of sense amplifiers and other matched circuits if matched components degrade differently [Huh98]. Hot electron degradation can be analyzed with simulators [Hu92, Hsu91, Quader94]. The wear is limited by setting maximum values on input risetime and stage electrical effort [Leblebici96]. These maximum values depend on the process and operating voltage. Weste04r4.fm Page 152 Monday, January 5, 2004 1:24 AM 152 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION A related aging mechanism is negative bias temperature instability (NBTI), which leads to a decrease in pMOS transistor current as transistors wear at high temperature [Doyle91]. NBTI results from trapped holes in the oxide coupled with the creation of interface states. Like hot carriers, it leads to circuit failures from increased delay and poorer matching [Reddy02]. NBTI shifts depend on the electric field seen by the device and can be locked in to the device by high-voltage stress during burn-in; this is good because it allows testing with full NBTI degradation. 4.8.5 Latchup Early adoption of CMOS processes was slowed by a curious tendency of CMOS chips to develop low-resistance paths between VDD and GND, causing catastrophic meltdown. The phenomenon, called latchup, occurs when parasitic bipolar transistors formed by the substrate, well, and diffusion turn ON. With process advances and proper layout procedures, latchup problems can be easily avoided. The cause of the latchup effect [Estreich82, Troutman86] can be understood by examining the process cross-section of a CMOS inverter, shown in Figure 4.63(a), over which is overlaid an equivalent circuit. In addition to the expected nMOS and pMOS transistors, the schematic depicts a circuit composed of an npn-transistor, a pnp-transistor, and two resistors connected between the power and ground rails (Figure 4.63(b)). The npn-transistor is formed between the grounded n-diffusion source of the nMOS transistor, the p-type substrate, and the n-well. The resistors are due to the resistance through the substrate or well to the nearest substrate and well taps. The cross-coupled transistors form a bistable silicon-controlled rectifier (SCR). Ordinarily, both parasitic bipolar transistors are OFF. Latchup can be triggered when transient currents flow through the substrate during normal chip power-up or when external voltages outside the normal operating range are applied. If substantial current flows in the substrate, Vsub will rise, turning ON the npn-transistor. This pulls current through the well resistor, bringing down Vwell and turning ON the pnp-transistor. The pnp-transistor current in turn raises Vsub, initiating a positive feedback loop with a large current flowing between VDD and GND that persists until the power supply is turned off or the power wires melt. Fortunately, latchup prevention is easily accomplished by minimizing Rsub and Rwell. Some processes use a thin epitaxial layer of lightly doped silicon on top of a heavily doped substrate that offers a low substrate resistance. Most importantly, the designer should place substrate and well taps close to each transistor, as described in Section 1.5.1. A conservative guideline is to place a tap adjacent to every source connected to VDD or GND. If this is not practical, you can obtain more detailed information from the process vendor (they will normally specify a maximum distance for diffusion to substrate/well tap) or try the following guidelines: Every well should have at least one tap. All substrate and well taps should connect directly to the appropriate supply in metal. Weste04r4.fm Page 153 Monday, January 5, 2004 1:24 AM 4.8 A GND V DD Y p+ n+ n+ p-substrate Rsub p+ p+ n+ Rwell n-well Vwell Vsub (a) Substrate Tap R well (b) Well Tap Vwell Vsub Rsub FIG 4.63 Origin and model of CMOS latchup A tap should be placed for every 5-10 transistors or every 25-100 µm (this distance is process-dependent). nMOS transistors should be clustered together near GND and pMOS transistors should be clustered together near VDD, avoiding convoluted structures that intertwine nMOS and pMOS transistors in checkerboard patterns. I/O pads are especially susceptible to latchup because external voltages can ring below GND or above VDD, forward biasing the junction between the drain and substrate or well and injecting current into the substrate. In such cases, guard rings should be used to collect the current, as shown in Figure 4.64. Guard rings are simply substrate or well taps tied to the proper supply that completely surround the transistor of concern. For example, the n+ diffusion in Figure 4.64(b) can inject electrons into the substrate if it falls a diode drop below 0 volts. The p+ guard ring tied to ground provides a low-resistance path to collect these electrons before they interfere with the operation of other circuits outside the guard ring. All diffusion structures in any circuit connected to the external world must be guard ringed, i.e., n+ diffusion by p+ connected to GND or p+ diffusion by n+ connected to VDD. RELIABILITY 153 Weste04r4.fm Page 154 Monday, January 5, 2004 1:24 AM 154 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION n well V DD p+ n+ GND n+ guard ring (a) p+ guard ring p substrate (b) FIG 4.64 Guard rings For the ultra-paranoid, double guard rings may be employed, i.e., n+ ringed by p+ to GND, then n+ to VDD or p+ ringed by n+ to VDD, then p+ to GND. SOI processes avoid latchup entirely because they have no parasitic bipolar structures. Also, processes with VDD < 0.7 are immune to latchup because the parasitic transistors will never have a large enough base-to-emitter voltage to turn on. In general, low-voltage processes are much less susceptible to latchup problems. 4.8.6 Overvoltage Failure Tiny transistors can be easily damaged by relatively low voltages. O vervoltage reliability problems can arise from electrostatic discharge, oxide breakown, punchthrough, and timedependent dielectric breakdown of the gate oxide. Electrostatic discharge (ESD) from static electricity entering the I/O pads can cause very large voltage and current transients and is discussed further in Section 12.4. Undesired voltages applied to the gate can cause breakdown and arcing across the thin oxide, destroying the device. Higher-than-normal voltages applied between source and drain lead to punchthrough when the source/drain depletion regions touch. Both problems lead to a maximum safe voltage that can be applied to transistors. For modern processes, this voltage is often much less than the I/O standard voltage, requiring a second type of transistor with thicker oxides and longer channels to endure the higher I/O voltages. Gate oxides wear out with time as tunneling currents cause irreversible damage to the oxide; this problem is called time-dependent dielectric breakdown (TDDB). The failure rate is exponentially dependent on the temperature and oxide thickness; for a 10-year life at 125° C, the field across the gate Eox = VDD/tox should be kept below about 7 MV/cm = 0.7 V/nm [Moazzami90]. The problem is greatest when voltage overshoots occur; this can Weste04r4.fm Page 155 Monday, January 5, 2004 1:24 AM 4.9 be caused by noisy power supplies or reflections at I/O pads. Reliability is improved by lowering the power supply voltage, minimizing power supply noise, and using thicker oxides on the I/O pads. 4.8.7 Soft Errors In the 1970s, as dynamic RAMs (DRAMs) replaced core memories, DRAM vendors were puzzled to find DRAM bits occasionally flipping value spontaneously. At first, the errors were attributed to “system noise,” “voltage marginality,” “sense amplifiers,” or “pattern sensitivity,” but the errors were found to be random. When the corrupted bit was rewritten with a new value, it was no more likely than any other bit to experience another error. In a classic paper [May79], Intel identified the source of these soft errors as alpha particle collisions that generate electron-hole pairs in the silicon as the particles lose energy. The excess carriers can be collected into the diffusion terminals of transistors. If the charge collected is comparable to the charge on the node, the voltage can be disturbed. Soft errors are random nonrecurring single bit errors in memory devices, including SRAM, DRAM, registers, and latches. Alpha particles from decaying uranium and thorium impurities in integrated circuit interconnect and packaging is a major source of soft errors at sea level, often causing a soft error rate (SER) of 100–2000 FIT/Mb [Hazucha00]. The neutron flux from cosmic rays is low at sea level, but two orders of magnitude larger at aircraft flight altitudes [Ziegler96]. These cosmic rays cause up to 106 FIT/Mb at flight altitudes. Soft errors are minimized by maintaining at least some critical charge Qcrit (Q = CV ) on state nodes. This is difficult as memory cells become smaller and hence have less capacitance, and as power supplies get lower. DRAMs use exotic structures such as trench capacitors to hold sufficient charge. Fortunately, small cells require less Qcrit because of their small area to collect carriers. The error rate is also affected by the proximity of alpha sources. Flip-chip technology with solder bumps bonded directly to the die poses special risks because impurities in the lead bumps emit a high alpha particle flux. Some companies avoid placing solder bumps directly over RAMs or other sensitive circuits. Aged lead (e.g., from Roman pipes) has outlived the contaminant half-lives and is much safer. Similarly, highly purified aluminum interconnect reduces the alpha flux from chip wires [ Juhnke95]. Finally, error detecting and correcting codes can be used to tolerate soft errors in memories without data corruption. These codes will be discussed further in Section 10.7.2. 4.9 Scaling The only constant in VLSI design is constant change. Figure 4.65 shows the year of introduction of Intel microprocessors in each feature size on a logarithmic scale, indicating that feature size reduces by 30% from one generation to the next every 2 to 3 years. As transis- SCALING 155 Weste04r4.fm Page 156 Monday, January 5, 2004 1:24 AM CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION tors become smaller, they switch faster, dissipate less power, and are cheaper to manufacture! Despite the ever-increasing challenges, process advances have actually accelerated in the past decade. Such scaling is unprecedented in the history of technology. However, scaling also exacerbates noise and reliability issues and introduces new problems. Designers need to be able to predict the effect of this feature size scaling on chip performance to plan future products, ensure existing products will scale gracefully to future processes for cost reduction, and anticipate looming design challenges. This section examines how transistors and interconnect scale, and the implications of scaling for design. The Semiconductor Industry Association prepares and maintains an International Technology Roadmap for Semiconductors predicting future scaling. Section 4.11 gives a case study of how scaling has influenced Intel microprocessors over three decades. 10 10 Feature Size (mm) 156 6 3 1 1.5 1 0.8 0.6 0.35 0.25 0.18 0.13 0.09 0.1 1965 1970 1975 1980 1985 1990 1995 2000 2005 Year FIG 4.65 Year of introduction of processes 4.9.1 Transistor Scaling First-order constant field MOS scaling theory is based on a model formulated by Dennard [Dennard74]. The characteristics of an MOS device can be maintained and the basic operational characteristics can be preserved if the critical parameters of a device are scaled by a dimensionless factor S. These parameters include All dimensions (in the x, y, and z directions) Device voltages Doping concentration densities Weste04r4.fm Page 157 Monday, January 5, 2004 1:24 AM 4.9 Another approach is lateral scaling, in which only the gate length is scaled. This is commonly called a gate shrink because it can be done easily to an existing mask database for a design. The effects of these types of scaling are illustrated in Table 4.15. The industry generally scales process generations with S = 2 . This doubles the number of transistors per unit area with each generation and doubles transistor performance every two generations under constant field scaling. A 5% gate shrink (S = 1.05) is commonly applied as a process becomes mature to boost the speed of components in that process. Table 4.15 Influence of scaling on MOS device characteristics Parameter Sensitivity Constant Field Lateral 1/S 1/S 1/S 1/S 1/S S 1/S 1 1 1 1 1 S S 1/S S 1 1/S 1/S 1/S 1/S S 1/S2 1/S2 1 S 1/S2 S2 S 1 S S Scaling Parameters Length: L Width: W Gate oxide thickness: tox Supply voltage: VDD Threshold voltage: Vtn, Vtp Substrate doping: NA Device Characteristics β Current: Ids W1 L tox β(V DD − V t ) Resistance: R V DD I ds Gate capacitance: C Gate delay: τ Clock frequency: f Dynamic power dissipation (per gate): P Chip area: A Power density Current density WL tox RC 1/τ CV2f P/A Ids/A 2 SCALING 157 Weste04r4.fm Page 158 Monday, January 5, 2004 1:24 AM 158 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION For constant field scaling, all device dimensions, including channel length L, width W, and oxide thickness tox, are reduced by a factor of 1/S. The supply voltage VDD and the threshold voltages are also reduced by 1/S. The substrate doping NA is increased by S. Because both distance and voltage are scaled equally, the electric field remains constant. This has the desirable effect that many nonlinear factors essentially remain unaffected. Because both L and tox scale at the same rate, the gate capacitance per micron of width Cpermicron has also remained approximately constant at 1.5–2 fF/µm, as described in Section 2.3.1. A gate shrink scales only the channel length, leaving other dimensions, voltages, and doping levels unchanged. This offers a quadratic improvement in gate delay according to the first order model. In practice, the gate delay improvement is closer to linear because velocity saturation keeps the current and effective resistance approximately constant. Historically, feature sizes were shrunk from 6 µm to 1 µm while maintaining a 5 V supply voltage. This was called c onstant voltage scaling and offered quadratic delay improvement as well as cost reduction. It also maintained continuity in I/O voltage standards. Constant voltage scaling increased the electric fields in devices. By the 1 µm generation, velocity saturation was severe enough that decreasing feature size no longer improved device current. Device breakdown from the high field was another risk. Therefore, constant field scaling has been the rule for modern devices. Example Most processes have gate capacitance of roughly 2 fF/µm. If the FO4 inverter delay of a process with features size f (in nm) is 1/2 ps • f, estimate the ON resistance of a unit (i.e., 4 λ wide) nMOS transistor. Solution: An FO4 inverter has a delay of 5 τ = 15 RC. Therefore, RC = 1 2 f 15 = f ps 30 nm (4.74) A unit transistor has width W = 2f and thus capacitance of C = 4f fF/µm. Solving for R, f ps 1 µm R= = 8.33 kΩ 30 nm 4 f fF (4.75) Note that this is independent of feature size. The resistance of a unit transistor is roughly independent of feature size, while the gate capacitance decreases with feature size. Alternatively, the capacitance per micron is roughly independent of feature size while the resistance • micron decreases with feature size. Weste04r4.fm Page 159 Monday, January 5, 2004 1:24 AM 4.9 The FO4 inverter delay will scale as 1/S assuming ideal constant-field scaling. As we saw in Section 4.7.4, this delay is commonly 1/2 ps/nm of the effective channel length for typical processing and worst-case environment. Aggressive processes achieve delays in the short end of the range by building transistors with effective channel lengths somewhat shorter than the feature size would imply. 4.9.2 Interconnect Scaling Two common approaches to interconnect scaling are to either scale all dimensions or keep the wire height constant. Table 4.16 shows the resistance, capacitance, and delay per unit length for each of these techniques. Wire length decreases for some types of wires, but may increase for others. Local and scaled wires are those that decrease in length during scaling. For example, a wire across a 64-bit ALU is local because it becomes shorter as the ALU is migrated to a finer process. A wire across a particular microprocessor is scaled because when the microprocessor is shrunk to the new process, the wire will also shrink. However, wires crossing next generation microprocessors are global because the next generation microprocessor die is likely to be larger (by a factor of Dc, on the order of 1.1) than the previous generation, so the cross-chip wires get longer rather than shorter. The table also shows scaling of wire delay for interconnect with and without repeaters. Unrepeated interconnect delay is remaining about constant for local interconnect and increasing for global interconnect. This presents a problem because transistors are getting faster, so the ratio of interconnect to gate delay increases with scaling. An increasing fraction of circuits are limited by wire delay rather than gate delay, and designers must devote greater attention to wire engineering and repeaters. In older processes where wire width and spacing were much greater than wire thickness, it was advantageous to scale wires by reducing the width and spacing, but not the thickness. This avoids the quadratic increase in wire resistance per unit length and was acceptable because fringing capacitance was a small fraction of the whole. In modern processes with aspect ratios of 1.5–2.2, fringing capacitance accounts for the majority of the total capacitance. Scaling spacing but not height increases the fringing capacitance enough that the extra thickness scarcely improves delay. Moreover, the coupling capacitance to nearby neighbors causes severe crosstalk. Therefore, it is now common to reduce thickness of lower-level metal interconnect with each generation. Of course, process engineers can choose from a continuum of possibilities between linearly scaling wire thickness and keeping wire thickness constant. Observe that when wire thickness is scaled, the capacitance per unit length remains constant. Hence, a reasonable initial estimate of the capacitance of a minimum-pitch wire is about 0.2 fF/µ m, independent of the process. In other words, wire capacitance is roughly 1/10–1/6 of gate capacitance per unit length. SCALING 159 Weste04r4.fm Page 160 Monday, January 5, 2004 1:24 AM 160 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Table 4.16 Influence of scaling on interconnect characteristics Parameter Sensitivity Reduced Thickness Constant Thickness Scaling Parameters Width: w Spacing: s Thickness: t Interlayer oxide height: h 1/S 1/S 1/S 1 1/i Characteristics Per Unit Length Water resistance per unit length: Rw 1 wt S2 S t s 1 S w h 1 1 Total wire capacitance per unit length: Cw Cwf + Cwp 1 between 1, S Unrepeated RC constant per unit length: twu Repeated wire RC delay per unit length: twr (assuming constant field scaling of gates in Table 4.15) Crosstalk noise RwCw S2 RCRwC w S between S, S2 between 1, S t s 1 Length: l Unrepeated wire RC delay l2twu 1 Repeated wire delay ltwr 1/ S Length: l Unrepeated wire RC delay l2twu S2Dc2 Repeated wire delay ltwr Fringing capacitance per unit length: Cwf Parallel plate capacitance per unit length: Cwp S Local/Scaled Interconnect Characteristics 1/S between 1/S, 1 between 1/S, 1 / S Global Interconnect Characteristics Dc S Dc between SDc2, S2Dc2 between Dc , Dc S Weste04r4.fm Page 161 Monday, January 5, 2004 1:24 AM 4.9 4.9.3 International Technology Roadmap for Semiconductors The incredible pace of scaling requires cooperation among many companies and researchers both to develop compatible process steps and to anticipate and address future challenges before they hold up production. The Semiconductor Industry Association develops and updates the International Technology Roadmap for Semiconductors [ITRS02] to forge a consensus so that development efforts are not wasted on incompatible technologies and to predict future needs and direct research efforts. Such an effort to predict the future is inevitably prone to error, and the industry has scaled feature sizes and clock frequencies more rapidly than the roadmap predicted in the late 1990s. Nevertheless, the roadmap offers a more coherent vision than one could obtain by simply interpolating straight lines through historical scaling data. The ITRS forecasts a major new technology generation, also called technology node, approximately every three years. The scaling between generations is traditionally S= 2 so the number of transistors per unit area doubles every generation. Table 4.17 summarizes some of the predictions, particularly for high-performance microprocessors. However, serious challenges lie ahead, and major breakthroughs will be necessary in many areas to maintain the scaling on the roadmap. The FO4 delays/cycle figure is extracted assuming the FO4 delay in picoseconds is one-third of the feature size in nanometers. These cycle times appear problematic because it is very difficult for a digital signal to swing rail to rail in less than 6 FO4 delays. Refer to the ITRS for more and annual updates. Table 4.17 Predictions from the 2002 ITRS Year Feature size (nm) VDD (V) Millions of transistors/die Wiring levels Intermediate wire pitch (nm) Interconnect dielectric constant I/O signals Clock rate (MHz) FO4 delays/cycle Maximum power (W) DRAM capacity (Gbits) 2001 130 2004 90 2007 65 2010 45 2013 32 2016 22 1.1–1.2 193 8–10 450 3–3.6 1–1.2 385 9–13 275 2.6–3.1 0.7–1.1 773 10–14 195 2.3–2.7 0.6–1.0 1564 10–14 135 2.1 0.5–0.9 3092 11–15 95 1.9 0.4–0.9 6184 11–15 65 1.8 1024 1684 13.7 130 0.5 1024 3990 8.4 160 1 1024 6739 6.8 190 4 1280 11511 5.8 218 8 1408 19348 4.8 251 32 1472 28751 4.7 288 64 SCALING 161 Weste04r4.fm Page 162 Monday, January 5, 2004 1:24 AM 162 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION 4.9.4 Impacts on Design One of the limitations of first-order scaling is that it gives the wrong impression of being able to scale proportionally to zero dimensions and zero voltage. In reality, a number of factors change significantly with scaling. This section attempts to peer into the crystal ball and predict some of the impacts on design for the future. These predictions are notoriously risky because chip designers have had an astonishing history of inventing ingenious solutions to seemingly insurmountable barriers. 4.9.4.1 Improved Performance and Cost The most positive impact of scaling is that performance and cost are steadily improving. System architects need to understand the scaling of CMOS technologies and predict the capabilities of the process several years into the future, when a chip will be completed. Because transistors are becoming cheaper each year, architects particularly need creative ideas of how to exploit growing numbers of transistors to deliver more or better functions. When transistors were first invented, the best predictions of the day suggested that they might eventually approach a fifty-cent manufacturing cost. Figure 4.66 plots the number of transistors and average price per transistor shipped by the semiconductor industry over the past three decades [Moore03]. In 2003, you could buy more than 100,000 transistors for a penny. FIG 4.66 Transistor shipments and average price 4.9.4.2 Interconnect Scaled transistors are steadily improving in delay, but scaled wires are holding constant or getting worse. Figure 4.67, taken from the 1997 Semiconductor Industry Association Roadmap [SIA97], showed the sum of gate and wire bottoming out at the 250 or 180 nm generation and getting worse thereafter. The wire problem motivated a number of papers predicting the demise of conventional wires. However, the plot Weste04r4.fm Page 163 Monday, January 5, 2004 1:24 AM 4.9 FIG 4.67 Gate and wire delay scaling is misleading in two ways. First, the “gate” delay is shown for a single unloaded transistor (delay = RC) rather than a realistically loaded gate (e.g., an FO4 inverter delay = 15RC). Second, the wire delays shown are for fixed lengths, but as technology scales, most local wires connecting gates within a unit also become shorter [Ho01]. In practice, for short wires, such as those inside a logic gate, the wire RC delay is negligible and will remain so for the foreseeable future. However, the long wires present a considerable challenge. It is no longer possible to send a signal from one side of a large, high-performance chip to another in a single cycle. Also, the “reachable radius” that a signal can travel in a cycle is steadily getting smaller, as shown in Figure 4.68. This requires that microarchitects understand the floorplan and budget multiple pipeline stages for data to travel long distances across the die. Repeaters help somewhat, but even so, interconnect does not keep up. Moreover, the “repeater farms” must be allocated space in the floorplan. As scaled gates become faster, the delay of a repeater goes down and hence, you should expect it will be better to use more repeaters. This means a greater number of repeater farms are required. One technique to alleviate the interconnect problem is to use more layers of interconnect. Table 4.18 shows the number of layers of interconnect increasing with each generation in TSMC processes. The lower layers of interconnect are classically scaled to provide high-density short connections. The higher layers are scaled less aggressively, or possibly even reverse-scaled to be thicker and wider to provide low-resistance, high-speed interconnect, good clock distribution networks, and a stiff power grid. Copper and low-k dielectrics were also introduced to reduce resistance and capacitance. SCALING 163 Weste04r4.fm Page 164 Monday, January 5, 2004 1:24 AM 164 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Chip Size Scaling of Reachable Radius FIG 4.68 Reachable radius scaling Table 4.18 Scaling of metal layers in TSMC processes Process (nm) Metal Layers 500 350 250 180 150 130 90 3 (Al) 4 (Al) 5 (Al) 6 (Al, low-k) 7 (Cu, low-k) 8 (Cu, low-k) 9 (Cu, low-k) Blocks of 50–100 Kgates (1 Kgate = 1000 3-input NAND gates or 6000 transistors) will continue to have reasonably short internal wires and acceptably low wire RC delay [Sylvester98]. Therefore, large systems can be partitioned into blocks of roughly this size with repeaters inserted as necessary for communication between blocks. 4.9.4.3 Power In classical constant field scaling, power density remains constant and overall chip power increases only slowly with die size. In practice, power density has skyrocketed because clock frequencies have increased much faster than classical scaling would predict and VDD is somewhat higher than constant field scaling would demand. Intel Vice President Patrick Gelsinger gave a keynote speech at the International Solid State Circuits Conference on February 5, 2001 [Gelsinger01]. He showed that microprocessor power consumption had been increasing exponentially (see Figure 4.69) Weste04r4.fm Page 165 Monday, January 5, 2004 1:24 AM 4.9 100000 Power (W) 10000 1000 100 Pentium® processors 286 486 8086 386 8085 8080 8008 1 8004 100 0.1 1971 1974 1978 1985 1992 2000 2004 2008 Year FIG 4.69 Intel processor power consumption and was predicted to grow even faster in the coming decade. He predicted “business as usual will not work in the future,” and that if scaling continued at this pace, by 2005, highspeed processors would have the power density of a nuclear reactor, by 2010, a rocket nozzle, and by 2015, the surface of the sun! The next day, Intel’s stock dropped 8% and Intel has since downplayed the forecast because designers will obviously devote more attention to controlling power. Dynamic power consumption will not continue to increase at such rates because it will become uneconomical to cool the chips. One reason is that at the time of this writing, high-performance processors are operating at about 12–16 FO4 inverter delays/cycle. It will be very difficult to generate a clock with a period of less than 6 FO4 inverter delays because the clock will look like a sinusoid rather than a square wave [Ho01] and because sequencing overhead becomes excessive; hence, frequency cannot scale faster than raw gate delays too much longer. Another reason is that cache area is likely to become a larger fraction of the die area and caches have low activity factors and lower power dissipation per unit area. Nevertheless, designers will need to budget and plan for power consumption as a factor nearly as important as performance and perhaps more important than area. Static power consumption will be a growing concern, especially for battery-operated devices. The static power consumption caused by subthreshold leakage was historically negligible but becomes important for threshold voltages below about 0.3–0.4 V, and may become comparable to dynamic power for systems operating below 1 V. Figure 4.70 shows how the static power has historically increased much more rapidly than dynamic power [Moore03]. This is especially problematic because simply turning off clocks in a sleep mode is not sufficient to stop the static power consumption. To slow the increase of leakage, Vt has changed from about VDD/5 in older processes to VDD/3 in newer processes as SCALING 165 Weste04r4.fm Page 166 Monday, January 5, 2004 1:24 AM CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION 1000 100 Power (W) 166 10 1 0.1 0.01 0.001 1960 1970 1980 1990 2000 2001 Year FIG 4.70 Dynamic and static power trends VDD has dropped; this has eliminated circuits with threshold drops from the set of viable circuit styles. Systems will use multiple flavors of transistors with different threshold voltages so that gates on the critical path can have fast low-Vt devices while memories and noncritical gates save power with higher-Vt devices. Gate tunneling current is also important for oxides of less than 15–20 Å. Multiple oxide thicknesses are also required. I/O standards of 1.8 or 3.3 V are destructive to the very thin oxides in modern processes, so thicker oxides are used for slower I/O transistors. Thicker oxides may also become an option for memories and noncritical circuits. Even if power remains constant, lower supply voltage leads to higher current density. This in turn causes higher IR drops and di/dt noise in the supply network (see Sections 6.3.5 and 12.3). These factors lead to more pins and metal resources on a chip being required for the power distribution networks. 4.9.4.4 Productivity The number of transistors that fit on a chip is increasing faster than designer productivity (gates/week). This leads to design teams of increasing size, and difficulty recruiting enough experienced engineers when the economy is good. It has driven a search for design methodologies that maximize productivity, even at the expense of performance and area. Now most chips are designed using synthesis and place and route; the number of situations where custom circuit design is affordable is diminishing. In other words, creativity is shifting from the circuit to the systems level for many designs. On the other hand, performance is still king in the microprocessor world. Design teams in that Weste04r4.fm Page 167 Monday, January 5, 2004 1:24 AM 4.9 field are approaching the size of automotive and aerospace teams because the development cost is justified by the size of the market. This drives a need for engineering managers who are skilled in leading such large organizations. The number of 50–100 Kgate blocks is growing, even in relatively low-end systems. This demands greater attention to floorplanning and placement of the blocks. One of the best hopes to solve the productivity gap is design reuse. Intellectual property (IP) blocks can be purchased and used as black boxes within a system-on-chip (SoC) in much the same way chips are purchased for a board-level design. 4.9.4.5 Physical Limits How far will CMOS processes scale? It is clear that scaling cannot continue indefinitely; transistors as we know them today will not work if the oxide is less than an atomic layer thick, the channel less than an atomic layer long, or the charge in the channel less than that of one electron. Numerous papers have been written forecasting the end of silicon scaling. For example, in 1972, the limit was placed at the 0.25 µm generation because of tunneling and fluctuations in dopant distributions [Hoeneisen72, Mead80]; at this generation, chips were predicted to operate at 10–30 MHz! In 1999, IBM predicted that scaling would nearly grind to a halt beyond the 100 nm generation in 2004 [Davari99]. In the authors’ experience, seemingly insurmountable barriers have seemed to loom about a decade away. Reasons given for these barriers have included: Dynamic power dissipation Subthreshold leakage at low VDD and Vt Tunneling current through thin oxides Poor I-V characteristics due to short channel effects Optics for cost-effective manufacturing of small features Exponentially increasing costs of fabrication facilities and mask sets Electromigration Interconnect delay At the time of writing, it appears that no fundamental barriers exist before the 35 nm generation in 2013 (roughly a decade ahead). Beyond this point, it is difficult to predict the future. Nevertheless, a large number of extremely talented people are continuously pushing the limits and hundreds of billions of dollars are at stake, so we are reluctant to bet against the future of scaling. SCALING 167 Weste04r4.fm Page 168 Monday, January 5, 2004 1:24 AM 168 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION 4.10 Pitfalls and Fallacies Defining gate delay for an unloaded gate When marketing a process, it is common to report gate delay based on an inverter in a ring oscillator (2τ) or even the RC time constant of a transistor charging its own gate capacitance (1/3 τ). Remember that the delay of a real gate on the critical path should be closer to 5– 6τ. When in doubt, ask how “gate delay” is defined or ask for the FO4 inverter delay. Trying to increase speed by increasing the size of transistors in a path Most designers know that increasing the size of a transistor decreases its resistance and thus makes it faster at driving a constant load. Novice designers sometimes forget that increasing the size increases input capacitance and makes the previous stage slower, especially when that previous stage belongs to somebody else’s timing budget. The authors have seen this lead to lack of convergence in full-chip timing analysis on a large microprocessor because individual engineers boost the size of their own gates until their path meets timing. Only after the weekly full-chip timing roll-up do they discover that their inputs now arrive later because of the greater load on the previous stage. The solution is to include in the specification of each block not only the arrival time but also the resistance of the driver in the previous block. Trying to increase speed by using as few stages of logic as possible Logic designers often count “gate delays” in a path. This is a convenient simplification when used properly. In the hands of an inexperienced engineer who believes each gate contributes a gate delay, it suggests that the delay of a path is minimized by using as few stages of logic as possible, which is clearly untrue. Designing a large chip without considering the floorplan In the mid-1990s, designers became accustomed to synthesizing a chip from HDL and “tossing the netlist over the wall” to the vendor who would place and route it and manufacture the chip. Many designers were shielded from considering the physical implementation. Now flight times across the chip are a large portion of the cycle time in slow systems and multiple cycles in faster systems. If the chip is synthesized without a floorplan, some paths with long wires will be discovered to be too slow after layout. This requires resynthesis with new timing constraints to shorten the wires. When the new layout is completed, the long wires simply show up in different paths. The solution to this convergence problem is to make a floorplan early and microarchitect around this floorplan, including budgets for wire flight time between blocks. Algorithms termed timing directed placement have alleviated this problem, resulting in place & route tools that converge in one or a few iterations. Not stating process corner or environment when citing circuit performance Most products must be guaranteed to work at high temperature, yet many papers are written with transistors operating at room temperature (or lower), giving optimistic performance results. For example, at the International Solid State Circuits Conference Intel described a Pentium II processor running at a surprisingly high clock rate [Choudhury97], but when asked, the speaker admitted that the measurements were taken while the processor was “colder than an ice cube.” Similarly, the FFFFF design corner is sometimes called the “published paper” corner because delays are reported under these simulation or manufacturing con- Weste04r4.fm Page 169 Monday, January 5, 2004 1:24 AM 4.11 ditions without bothering to state that fact or report the FO4 inverter delay in the same conditions. Circuits in this corner are about twice as fast as in a manufacturable part. Providing too little margin in matched delays We have seen that the delay of a chain of inverters can vary by about 30% as compared to the delay of other circuits across design corners, voltage, and temperature. On top of this, you should expect intra-die process variation and errors in modeling and extraction. If a race condition exists where the circuit will fail when the inverter delay is faster than the gate delay, the experienced designer who wishes to sleep well at night provides generous delay margin under nominal conditions. Remember that the consequences of too little margin can be a million dollars in mask costs for another revision of the chip and far more money in the opportunity cost of arriving late to market. HISTORICAL PERSPECTIVE 169 originally designed and manufactured on a 0.6 micron BiCMOS process. The Pentium II is a closely related derivative manufactured in a 0.35 micron process operating at a lower voltage. In the new process, bipolar transistors ceased to offer performance advantages and were removed at considerable design effort. Further derivatives of the same architecture migrated to 0.25 and 0.18 micron processes in which wire delay did not improve at the same rate as gate delay. Interconnect-dominated paths required further redesign to achieve good performance in the new processes. In contrast, the Pentium 4 was designed with process scaling in mind. Knowing that over the lifetime of the product, device performance would improve but wires would not, designers overengineered the interconnect-dominated paths for the original process so that the paths would not limit performance improvement as the process advanced [Deleganes02]. Failing to plan for process scaling Many products will migrate through multiple process generations. For example, the Intel Pentium Pro was 4.11 Historical Perspective The incredible history of scaling can be seen in the advancement of the microprocessor. The Intel microprocessor line spans more than three decades. Table 4.19 summarizes the progression from the first 4-bit microprocessor, the 4004, through the Pentium 4, courtesy of the Intel Museum [Intel03]. Over the three decades, feature size has improved nearly one hundred-fold. Transistor budgets multiplied by more than 10,000. Even more remarkably, clock frequencies have also multiplied by almost as much. Even as the challenges have grown in the past decade, scaling has accelerated. Weste04r4.fm Page 170 Monday, January 5, 2004 1:24 AM 170 CHAPTER 4 Table 4.19 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION History of Intel microprocessors over three decades Processor Year Feature Size (µm) Transistors Frequency (MHz) Word size Package 4004 8008 8080 8086 80286 Intel386 Intel486 Pentium Pentium Pro Pentium II Pentium III Pentium 4 1971 1972 1974 1978 1982 1985 1989 1993 1995 1997 1999 2001 10 10 6 3 1.5 1.5–1.0 1–0.6 0.8–0.35 0.6–0.35 0.35–0.25 0.25–0.18 0.18–0.13 2.3k 3.5k 6k 29k 134k 275k 1.2M 3.2–4.5M 5.5M 7.5M 9.5–28M 42–55M 0.75 0.5–0.8 2 5–10 6–12 16–25 25–100 60–300 166–200 233–450 450–1000 1400–3200 4 8 8 16 16 32 32 32 32 32 32 32 16-pin DIP 18-pin DIP 40-pin DIP 40-pin DIP 68-pin PGA 100-pin PGA 168-pin PGA 296-pin PGA 387-pin MCM PGA 242-pin SECC 330-pin SECC2 478-pin PGA Die photos of the microprocessors also illustrate the remarkable story of scaling. The 4004 [Faggin96] in Figure 4.71 was handcrafted to pack the transistors onto the tiny die. Observe the 4-bit datapaths and register files. Only a single layer of metal was available, so polysilicon jumpers were required when traces had to cross without touching. The masks were designed with colored pencils and were hand-cut from red plastic rubylith. Observe that diagonal lines were used routinely. The 16 I/O pads and bond wires are clearly visible. The processor was used in the Busicom calculator. The 80286 [Childs84] in Figure 4.72 shows a far more regular appearance. It is partitioned into regular datapaths, random control logic, and several arrays. The arrays include the instruction decoder PLA and memory management hardware. At this scale, individual transistors are no longer visible. The Intel486™ (originally 80486, but changed because a number cannot be trademarked) integrated an 8KB cache and floating point unit with a pipelined integer datapath, as shown in Figure 4.73. At this scale, individual gates are not visible. The center column is the 32-bit integer datapath. To its left is the cache, divided into four 2KB subarrays. Observe that the cache involves a significant amount of logic beside the subarrays. To the right are several blocks of synthesized control logic generated with automatic place and route tools. The “more advanced” tools no longer support diagonal interconnect. The wide datapaths in the lower left form the floating point unit. The Pentium Processor™ in Figure 4.74 provides a superscalar integer execution unit and separate 8KB data and instruction caches. The 32-bit datapath and its associated control logic is again visible in the center of the chip, although at this scale, the individual bitslices of the datapath are difficult to resolve. The instruction cache in the upper left Weste04r4.fm Page 171 Monday, January 5, 2004 1:24 AM 4.11 FIG 4.71 4004 microprocessor FIG 4.72 80286 microprocessor HISTORICAL PERSPECTIVE 171 Weste04r4.fm Page 172 Monday, January 5, 2004 1:24 AM 172 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION FIG 4.73 Intel486™ microprocessor feeds the instruction fetch and decode units to its right. The data cache is in the lower left. The bus interface logic sits between the two caches. The pipelined floating point unit, home of the infamous FDIV bug [Price95], is in the lower right. This floorplan is important to minimize wire lengths between units that often communicate, such as the instruction cache and instruction fetch or the data cache and integer datapath. The integer datapath often forms the heart of a microprocessor, and other units surround the datapath to feed it the prodigious quantities of instructions and data that it consumes. The Pentium™ III Processor, shown in Figure 4.75, offers out-of-order issue of up to three instructions per cycle. The entire left portion of the die is dedicated to 256–512 KB of level 2 cache to supplement the 32KB instruction and data caches. As processor performance outstrips memory system bandwidth, the portion of the die devoted to the cache hierarchy will continue to grow. The Pentium™ 4 Processor is shown in Figure 4.76. The complexity of a VLSI system is clear from the enormous number of separate blocks that were each uniquely designed by a team of engineers. Indeed, at this scale even major functional units become difficult to resolve. The high operating frequency is achieved with a long pipeline using about 14 FO4 inverter delays per cycle. Remarkably, portions of the integer execution unit are “double-pumped” at twice the regular chip frequency. Weste04r4.fm Page 173 Monday, January 5, 2004 1:24 AM 4.11 FIG 4.74 Pentium™ microprocessor FIG 4.75 Pentium™ III microprocessor HISTORICAL PERSPECTIVE 173 Weste04r4.fm Page 174 Monday, January 5, 2004 1:24 AM 174 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION FIG 4.76 Pentium™ 4 microprocessor Summary The VLSI designer’s challenge is to engineer a system that meets speed requirements while consuming little power or area, operating reliably, and taking little time to design. Circuit simulation is an important tool for calculating delay and will be discussed in depth in Chapter 5, but it takes too long to simulate every possible design; is prone to garbagein, garbage-out mistakes; and doesn’t give insight into why a circuit has a particular delay or how the circuit should be changed to improve delay. The designer must also have simple models to quickly estimate performance by hand and explain why some circuits are better than others. Although transistors are complicated devices with nonlinear current-voltage and capacitance-voltage relationships, for the purpose of delay estimation in digital circuits, they can be approximated quite well as having constant capacitance and an effective resistance R when ON. Logic gates are thus modeled as RC networks. The Elmore delay model estimates the delay of the network as the sum of each capacitance times the resistance through which it must be charged or discharged. Therefore, the gate delay consists of a parasitic delay (accounting for the gate driving its own internal parasitic capacitance) plus an effort delay (accounting for the gate driving an external load). The effort delay depends on the electrical effort (the ratio of load capacitance to input capacitance, also Weste04r4.fm Page 175 Monday, January 5, 2004 1:24 AM SUMMARY called fanout) and the logical effort (which characterizes the current driving capability of the gate relative to an inverter with equal input capacitance). Even in advanced fabrication processes, the delay vs. electrical effort curve fits a straight line very well. The method of Logical Effort builds on this linear delay model to help us quickly estimate the delay of entire paths based on the effort and parasitic delay of the path. We will use Logical Effort in subsequent chapters to explain what makes circuits fast. The power consumption of a circuit has both dynamic and static components. The dynamic power comes from charging and discharging the load capacitances and depends on the frequency, voltage, capacitance, and activity factor. The static power comes from circuits that have an intentional path from VDD to GND (like pseduo-nMOS) and from leakage. CMOS circuits have historically consumed relatively low power because complementary CMOS gates dissipate almost zero static power. However, leakage is increasing as feature size decreases, making static power consumption more important. As feature size decreases, transistors get faster but wires do not. Interconnect delays are now very important. The delay is again estimated using the Elmore delay model based on the resistance and capacitance of the wire and its driver and load. The wire delay grows with the square of its length, so long wires are often broken into shorter segments driven by repeaters. Vast numbers of wires are required to connect all the transistors, so processes provide many layers of interconnect packed closely together. The capacitive coupling between these tightly packed wires can be a major source of noise in a system. For reliable circuit operation, the designer must ensure that the circuit performs correctly across variations in the operating voltage and temperature, and must develop circuits that are robust for variations in transistor characteristics such as channel length or threshold voltage. Process corners are used to describe the worst-case combination of processing and environment for delay, power consumption, and functionality. The circuits must also be designed to operate correctly even as they age or are subject to cosmic rays and electrostatic discharge. CMOS processes have been steadily improving for more than 20 years and will continue to do so for at least the next decade. A good designer not only should be familiar with the capabilities of current processes, but also should be able to predict the capabilities of future processes as feature sizes get progressively smaller. In modern constant-field scaling, gate delay improves with channel length. The number of transistors on a chip grows quadratically. The energy for each transistor to switch decreases with the cube of channel length, but the dynamic power density remains about the same because chips have more transistors switching at higher rates. Static power goes up as small transistors have exponentially more leakage. Interconnect capacitance per unit length remains constant, but resistance increases because the wires have a smaller cross-section. Local wires get shorter and have constant delay, while global wires have increasing delay. Architects planning future chips will take advantage of the larger number of faster transistors, but must also rethink existing architectures because of the changing ratio of wire-to-gate delay and the increasing leakage current. 175 Weste04r4.fm Page 176 Monday, January 5, 2004 1:24 AM 176 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION Exercises 4.1 Sketch a 2-input NOR gate with transistor widths chosen to achieve effective rise and fall resistance equal to a unit inverter. Compute the rising and falling propagation delays of the NOR gate driving h identical NOR gates using the Elmore delay model. Assume that every source or drain has fully contacted diffusion when making your estimate of capacitance. 4.2 Sketch a stick diagram for the 2-input NOR. Repeat Exercise 4.1 with better capacitance estimates. In particular, if a diffusion node is shared between two parallel transistors, only budget its capacitance once. If a diffusion node is between two series transistors and requires no contacts, only budget half the capacitance because of the smaller diffusion area. 4.3 Find the rising and falling propagation delays of an AND-OR-INVERT gate using the Elmore delay model. Estimate the diffusion capacitance based on a stick diagram of the layout. 4.4 Find the worst-case Elmore delay of an n-input NOR gate. 4.5 Sketch a delay vs. electrical effort graph like that of Figure 4.8 for a 2-input NOR gate using the logical effort and parasitic delay estimated in Section 4.2.3. How does the slope of your graph compare to that of the 2-input NAND? How does the yintercept compare? 4.6 Let a 4x inverter have transistors four times as wide as those of a unit inverter. If a unit inverter has three units of input capacitance and parasitic delay of pinv, what is the input capacitance of a 4x inverter? What is the logical effort? What is the parasitic delay? 4.7 A three-stage logic path is designed so that the effort borne by each stage is 12, 6, and 9 delay units, respectively. Can this design be improved? Why? What is the best number of stages for this path? What changes do you recommend to the existing design? 4.8 Suppose a unit inverter with three units of input capacitance has unit drive. a) What is the drive of a 4x inverter? b) What is the drive of a 2-input NAND gate with 3 units of input capacitance? 4.9 Sketch a 4-input NAND gate with transistor widths chosen to achieve equal rise and fall resistance as a unit inverter. Show why the logical effort is 6/3. 4.10 Consider the two designs of a 2-input AND gate shown in Figure 4.77. Give an intuitive argument about which will be faster. Back up your argument with a calculation of the path effort, delay, and input capacitances x and y to achieve this delay. Weste04r4.fm Page 177 Monday, January 5, 2004 1:24 AM EXERCISES C C x y 6C (a ) C 6C (b) FIG 4.77 2-input AND gate 4.11 Consider four designs of a 6-input AND gate shown in Figure 4.78. Develop an expression for the delay of each path if the path electrical effort is H. What design is fastest for H = 1? For H = 5? For H = 20? (a ) (b) (c) (d) FIG 4.78 6-input AND gate 4.12 Repeat the decoder design example from Section 4.3.4 for a 32-word register file with 64-bit registers. Determine the fastest decoder design and estimate the delay of the decoder and the transistor widths to achieve this delay. 4.13 Design a circuit at the gate level to compute the following function: if (a == b) y = a; else y = 0; Let a, b, and y be 16-bit busses. Assume the input and output capacitances are each 10 units. Your goal is to make the circuit as fast as possible. Estimate the delay in FO4 inverter delays using Logical Effort if the best gate sizes were used. What sizes do you need to use to achieve this delay? 4.14 Plot the average delay from input A of an FO3 NAND2 gate from the datasheet in Figure 4.25. Why is the delay larger for the XL drive strength than for the other drive strengths? 4.15 Figure 4.79 shows a datasheet for a 2-input NOR gate in the Artisan Components standard cell library for the TSMC 180 nm process. Find the average parasitic delay and logical effort of the X1 NOR gate A input. Use the value of τ from Section 4.3.7. 177 Weste04r4.fm Page 178 Monday, January 5, 2004 1:24 AM 178 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION FIG 4.79 2-input NOR datasheet courtesy Artisan Components 4.16 Find the parasitic delay and logical effort of the X2 and X4 NOR gate A input. By what percentage do they differ from that of the X1 gate? What does this imply about our model that parasitic delay and logical effort depend only on gate type and not on transistor sizes? 4.17 What are the parasitic delay and logical effort of the X1 NOR gate B input? How and why do they differ from the A input? 4.18 Parasitic delay estimates in Section 4.2.4 are made assuming contacted diffusion on each transistor on the output node and ignoring internal diffusion. Would parasitic delay increase or decrease if you took into account that some parallel transistors on the output node share a single diffusion contact? If you counted internal diffusion capacitance between series transistors? If you counted wire capacitance within the cell? Weste04r4.fm Page 179 Monday, January 5, 2004 1:24 AM 179 EXERCISES 4.19 Consider a process in which pMOS transistors have three times the effective resistance as nMOS transistors. A unit inverter with equal rising and falling delays in this process is shown in Figure 4.80. Calculate the logical efforts of a 2-input NAND gate and a 2-input NOR gate if they are designed with equal rising and falling delays. 4.20 Generalize Exercise 4.19 if the pMOS transistors have µ times the effective resistance of nMOS transistors. Find a general expression for the logical efforts of a kinput NAND gate and a k-input NOR gate. As µ increases, comment on the relative desirability of NANDs vs. NORs. 3 A Y 1 FIG 4.80 Unit inverter 4.21 Some designers define a “gate delay” to be a fanout-of-3 2-input NAND gate rather than a fanout-of-4 inverter. Using Logical Effort, estimate the delay of a fanout-of3 2-input NAND gate. Express your result both in τ and in FO4 inverter delays, assuming pinv = 1. 4.22 Repeat Exercise 4.21 in a process with a lower ratio of diffusion to gate capacitance in which pinv = 0.75. By what percentage does this change the NAND gate delay, as measured in FO4 inverter delays? What if pinv = 1.25? 4.23 The 64-bit Naffziger adder [Naffziger96] has a delay of 930 ps in a fast 0.5-µm Hewlett-Packard process with an FO4 inverter delay of about 140 ps. Estimate its delay in a 70 nm process with an FO4 inverter delay of 20 ps. 4.24 An output pad contains a chain of successively larger inverters to drive the (relatively) enormous off-chip capacitance. If the first inverter in the chain has an input capacitance of 20 fF and the off-chip load is 10 pF, how many inverters should be used to drive the load with least delay? Estimate this delay, expressed in FO4 inverter delays. 4.25 The clock buffer in Figure 4.81 can present a maximum input capacitance of 100 fF. Both true and complementary outputs must drive loads of 300 pF. Compute the input capacitance of each inverter to minimize the worst-case delay from input to either output. What is this delay, in τ? Assume the inverter parasitic delay is 1. clk gclk clk FIG 4.81 Clock buffer 4.26 The clock buffer from Exercise 4.25 is an example of a 1-2 fork. In general, if a 1–2 fork has a maximum input capacitance of C1 and each of the two legs drives a load of C2, what should the capacitance of each inverter be and how fast will the circuit operate? Express your answer in terms of pinv. 4.27 A 180 nm standard cell process can have an average switching capacitance of 150 pF/mm2. You are synthesizing a chip composed of random logic with an average activity factor of 0.1. Estimate the power consumption of your chip if it has an area of 70 mm2 and runs at 450 MHz at VDD = 0.9 V. 4.28 You are considering lowering VDD to try to save power in a static CMOS gate. You will also scale Vt proportionally. Will dynamic power consumption go up or down? Will static power consumption go up or down? Weste04r4.fm Page 180 Monday, January 5, 2004 1:24 AM 180 CHAPTER 4 CIRCUIT CHARACTERIZATION AND PERFORMANCE ESTIMATION 4.29 Evaluate the benefits of the stack effect for subthreshold leakage by comparing I1 and I2 in Figure 4.29 (this problem is inspired by [Narendra01]). Assume all three transistors are identical and γ = 0, n = 1. a) If the transistors suffer from no DIBL (η = 0), prove that I2 /I1 = 1/2, just as you would expect if the transistors behaved as resistors. b) Does increasing η increase or decrease I1? By what fraction does I1 change at room temperature if η = 0.05 and VDD = 1.8 V? c) Does increasing η increase or decrease I2? By what fraction does I2 change at room temperature if η = 0.05 and VDD = 1.8 V? d) Solve for I2 /I1 and x as a function of ∆ = ηVDD/vT assuming ∆ >> 1. e) Explain why the stack effect is most important for transistors with significant DIBL. 4.30 Consider a 5 mm long, 4 λ-wide metal2 wire in a 0.6 µm process. The sheet resistance is 0.08 Ω/? and the capacitance is 0.2 fF/µm. Construct a 3-segment π-model for the wire. 4.31 A 10x unit-sized inverter drives a 2x inverter at the end of the 5 mm wire from Exercise 30. The gate capacitance is C = 2 fF/µm and the effective resistance is R = 2.5 kΩ • µm for nMOS transistors. Estimate the propagation delay using the Elmore delay model; neglect diffusion capacitance. 4.32 Find the best width and spacing to minimize the RC delay of a metal2 bus in the 180 nm process described in Table 4.8 if the pitch cannot exceed 1000 nm. Minimum width and spacing are 320 nm. First assume that neither adjacent bit is switching. How does your answer change if the adjacent bits can be switched? 4.33 Derive EQ (4.55)–(4.57). Assume the initial driver and final receiver are of the same size as the repeaters so the total delay is N times the delay of a segment. Neglect diffusion parasitics so each segment can be modeled as in Figure 4.82. R /W R wl/N C w l/N C'W FIG 4.82 Model of repeater driving interconnect and next gate Weste04r4.fm Page 181 Monday, January 5, 2004 1:24 AM EXERCISES 4.34 Revisit Exercise 33 using pair of inverters (a noninverting buffer) instead of a single inverter. The first inverter in each pair is W1 times unit width. The second is a factor of k larger than the first. Derive EQ (4.58)-(4.60). 4.35 Compute the characteristic velocity (delay per mm) of a repeated wire in the 180 nm process. A unit nMOS transistor has resistance of 2.5 KΩ and capacitance of 0.7 fF, and the pMOS has twice the resistance. Develop a table of results for metal1, metal2, and metal4 on minimum pitch and on double-pitch (twice minimum width and spacing). Use the data and pitches for the Intel 180 nm process. Assume solid metal above and below the wires and that the neighbors are not switching. 4.36 The Pentium 4 operated at 1.4 GHz in Intel’s 180 nm process with an FO4 delay of 50 ps in 2001. If a new process is introduced every two years with a scaling factor of S = 2, predict microprocessor clock frequencies in 2009, assuming all the improvement comes from ideal constant field scaling and that transistor speed sets clock rate. In practice, the frequency is likely to be higher. Why? 4.37 The path from the data cache to the register file of a microprocessor involves 500 ps of gate delay and 500 ps of wire delay along a repeated wire. The chip is scaled using constant field scaling and reduced height wires to a new generation with S = 2. Estimate the gate and wire delays of the path. By how much did the overall delay improve? 181 ...
View Full Document

This note was uploaded on 01/21/2010 for the course ECE260A 660090 taught by Professor Bendak,michaelbeshara during the Fall '09 term at UCSD.

Ask a homework question - tutors are online