DerekChiou_EE360N_Spring2010_Lecture23

DerekChiou_EE360N_Spring2010_Lecture23 - Lecture 23: Vector...

Info iconThis preview shows pages 1–8. Sign up to view the full content.

View Full Document Right Arrow Icon
ecture 3: ector Machines & Lecture 23: Vector Machines & Performance Evaluation Prof. Derek Chiou University of Texas at Austin © Derek Chiou 1
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Test of size nnouncements Announcements z I don’t have office hours tomorrow (Tuesday) but do today xam 2 on Wednesday z Exam 2 on Wednesday z Remember to bring all of your PDF lecture slides z Will be referring to them in the exam z Exam2 could potentially cover z Interrupts/exceptions z Buses z Pipelining z Delay slots z Scoreboarding z Tomasulo z ROB ranch prediction z Branch prediction z Vector machines 4/19/2010 © Derek Chiou: EE360N: Lecture 23 2
Background image of page 2
Test of size ecap and Outline Recap and Outline z Recap inish up branch prediction z Finish up branch prediction z Started on vector machines z Outline z Finish up vector machines z Performance measurement 4/19/2010 © Derek Chiou: EE360N: Lecture 23 3
Background image of page 3

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Test of size ector Machine: Performance Basics Vector Machine: Performance Basics z The contributions to run time of an instruction are tartup me (depth of the pipeline) z startup-time : (depth of the pipeline) z initiation rate : rate at which instructions enters the pipeline. z Under simplistic assumptions: z The time to complete a vector operation of length n is z startup-time + n / initiation rate 4/19/2010 © Derek Chiou: EE360N: Lecture 23 4
Background image of page 4
Test of size emory Initiation Rate Memory Initiation Rate z Memory speed must keep up with CPU speed! terleave memory banks z Interleave memory banks z To maintain initiation rate of 1, the number of banks ust be at least the ratio of memory to processor cycle must be at least the ratio of memory to processor cycle time. z However, if processor make access in non-unit stride, it is possible to have bank conflicts. 4/19/2010 © Derek Chiou: EE360N: Lecture 23 5
Background image of page 5

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Test of size emory Initiation Rate Memory Initiation Rate z In general, effective bandwidth (relative to 1 bank) as a function of stride and number of banks is??? performance- inded programmer will rearrange data to avoid z A performance minded programmer will rearrange data to avoid bad strides (by padding etc) z Alternatively, prime number of banks is an hardware solution. nly have most serious problem when accessing at stride z Only have most serious problem when accessing at stride exactly equal to the number of banks z Difficult to implement though. z ust do real divisions instead of shifts 4/19/2010 © Derek Chiou: EE360N: Lecture 23 6 Must do real divisions instead of shifts
Background image of page 6
Test of size emory Bandwidth Needed Memory Bandwidth Needed z Pipelined vector add z read two arguments, write one result back each cycle z 3 memory references / cycle z Cray-1, cycle time is 12.5ns (80MHz) z 3W * 80MHz = 240 M Words / sec z each word, 64b z 8B * 240M/sec ~= 2GB/sec (Cray-1 actually higher at 320MW/sec) xtremely fast for the time, not too bad even today z Extremely fast for the time, not too bad even today z Pentium 4 (a few years ago), cycle time is 0.2778ns (3.6GHz) z 3W * 3.6GHz = 10.8GW/sec ach word 64b z Each word, 64b z 8B * 10.8G/sec = 86.4GB/sec z Pentium 4 has 800Mb/sec * 64b bus = 6.4GB/sec peak (never chievable) 4/19/2010 © Derek Chiou: EE360N: Lecture 23 8 achievable) z Not very balanced
Background image of page 7

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Image of page 8
This is the end of the preview. Sign up to access the rest of the document.

This note was uploaded on 09/23/2010 for the course EE 360n taught by Professor Staff during the Spring '08 term at University of Texas at Austin.

Page1 / 35

DerekChiou_EE360N_Spring2010_Lecture23 - Lecture 23: Vector...

This preview shows document pages 1 - 8. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online