Computer Architecture Performance and Energy
0 likes•352 views
The document discusses computer architecture performance and energy consumption. It defines key performance metrics like response time and throughput. It also explains how to measure and calculate performance, including instruction count, clock cycles per instruction (CPI), and clock rate. The document emphasizes that performance depends on algorithms, programming languages, instruction sets, and technology. It discusses standardized benchmarks like SPEC CPU that are used to evaluate and compare overall processor performance using metrics like the geometric mean and geometric standard deviation of performance ratios.
![25
How Summarize Suite Performance (4/4)
n Does a single mean well summarize performance of programs in
benchmark suite?
n Can decide if mean a good predictor by characterizing variability
of distribution using standard deviation
n Like geometric mean, geometric standard deviation is multiplicative
rather than arithmetic
n Can simply take the logarithm of SPECRatios, compute the standard
mean and standard deviation, and then take the exponent to convert
back:
n The geometric standard deviation, denoted by σg, is calculated as
follows: log σg=[1/n∑n
i=1(logxi−logG)2]1/2.
n where G=n√x1⋅x2⋅…⋅xn is the geometric mean of SPECRatios (x1 . xn).
( )
( )
( )
( )
i
n
i
i
SPECRatio
StDev
tDev
GeometricS
SPECRatio
n
ean
GeometricM
ln
exp
ln
1
exp
1
=
÷
ø
ö
ç
è
æ
´
= å
=
CA: Performance and Power](https://image.slidesharecdn.com/computerarchitectureperformanceandenergy-240127061813-86754b21/85/Computer-Architecture-Performance-and-Energy-25-320.jpg)
![26
Example Standard Deviation: (1/3)
0
2000
4000
6000
8000
10000
12000
14000
wupwise
swim
mgrid
applu
mesa
galgel
art
equake
facerec
ammp
lucas
fma3d
sixtrack
apsi
SPECfpRatio
1372
5362
2712
GM = 2712
GStDev = 1.98
• GM and multiplicative StDev of SPECfp2000 for Itanium 2
Outside 1 StDev
Itanium 2 is
2712/100 times
as fast as Sun
Ultra 5 (GM), &
range within 1
Std. Deviation is
[13.72, 53.62]
CA: Performance and Power](https://image.slidesharecdn.com/computerarchitectureperformanceandenergy-240127061813-86754b21/85/Computer-Architecture-Performance-and-Energy-26-320.jpg)
![27
Example Standard Deviation : (2/3)
• GM and multiplicative StDev of SPECfp2000 for AMD Athlon
0
2000
4000
6000
8000
10000
12000
14000
wupwise
swim
mgrid
applu
mesa
galgel
art
equake
facerec
ammp
lucas
fma3d
sixtrack
apsi
SPECfpRatio
1494
2911
2086
GM = 2086
GStDev = 1.40
Outside 1 StDev
Athon is
2086/100 times
as fast as Sun
Ultra 5 (GM), &
range within 1
Std. Deviation is
[14.94, 29.11]
CA: Performance and Power](https://image.slidesharecdn.com/computerarchitectureperformanceandenergy-240127061813-86754b21/85/Computer-Architecture-Performance-and-Energy-27-320.jpg)
![28
Example Standard Deviation (3/3)
-
0.50
1.00
1.50
2.00
2.50
3.00
3.50
4.00
4.50
5.00
wupwise
swim
mgrid
applu
mesa
galgel
art
equake
facerec
ammp
lucas
fma3d
sixtrack
apsi
Ratio
Itanium
2
v.
Athlon
for
SPECfp2000
0.75
2.27
1.30
GM = 1.30
GStDev = 1.74
• GM and StDev Itanium 2 v Athlon
Outside 1 StDev
Ratio execution times (At/It) =
Ratio of SPECratios (It/At)
Itanium 2 1.30X Athlon (GM),
1 St.Dev. Range [0.75,2.27]
CA: Performance and Power](https://image.slidesharecdn.com/computerarchitectureperformanceandenergy-240127061813-86754b21/85/Computer-Architecture-Performance-and-Energy-28-320.jpg)
![35
Parallel Speed-up Examples (2/2)
n Consider summing 10 scalar variables and two
10 by 10 matrices (matrix sum) on 10
processors
Speedup = 1/(.091 + .909/10) = 1/0.1819 = 5.5
n What if there are 100 processors ?
Speedup = 1/(.091 + .909/100) = 1/0.10009 = 10.0
n What if the matrices are 100 by 100 (or 10,010
adds in total) on 10 processors?
Speedup = 1/(.001 + .999/10) = 1/0.1009 = 9.9
n What if there are 100 processors ?
Speedup = 1/(.001 + .999/100) = 1/0.01099 = 91
Speedup w/ E = 1 / [ (1-F) + F/S ]
CA: Performance and Power](https://image.slidesharecdn.com/computerarchitectureperformanceandenergy-240127061813-86754b21/85/Computer-Architecture-Performance-and-Energy-35-320.jpg)
Computer Architecture Performance and Energy
- 1. 1 Computer Architecture Performance and Energy Consumption Computer Architecture CA: Performance and Power
- 2. 2 Defining Performance n Which airplane has the best performance? CA: Performance and Power
- 3. 3 Response Time and Throughput n Response time n How long it takes to do a task n Throughput n Total work done per unit time n e.g., tasks/transactions/… per hour n How are response time and throughput affected by n Replacing the processor with a faster version? n Adding more processors? n We’ll focus on response time for now… CA: Performance and Power
- 4. 4 Relative Performance n Define Performance = 1/Execution Time n “X is n time faster than Y” n = = X Y Y X time Execution time Execution e Performanc e Performanc n Example: time taken to run a program n 10s on A, 15s on B n Execution TimeB / Execution TimeA = 15s / 10s = 1.5 n So A is 1.5 times faster than B CA: Performance and Power
- 5. 5 Measuring Execution Time n Elapsed time n Total response time, including all aspects n Processing, I/O, OS overhead, idle time n Determines system performance n CPU time n Time spent processing a given job n Discounts I/O time, other jobs’ shares n Comprises user CPU time and system CPU time n Different programs are affected differently by CPU and system performance CA: Performance and Power
- 6. 6 Analyze the Right Measurement! CA: Performance and Power Guides CPU design Guides system design CPU Time: Measuring CPU time (Ubuntu): $ time <program name> real 0m0.095s user 0m0.013s sys 0m0.008s Real elapsed time (in minutes and seconds) Time the CPU spends running program under measurement Total response time = CPU time + time spent waiting (for disk, I/O, ..)
- 7. 7 CPU Clocking n Operation of digital hardware governed by a constant-rate clock Clock (cycles) Data transfer and computation Update state Clock period n Clock period: duration of a clock cycle n e.g., 250ps = 0.25ns = 250×10–12 s n Clock frequency (rate): cycles per second n e.g., 4.0GHz = 4000MHz = 4.0×109 Hz CA: Performance and Power
- 8. 8 CPU Time n Performance improved by n Reducing number of clock cycles n Increasing clock rate n Hardware designer must often trade off clock rate against cycle count Rate Clock Cycles Clock CPU Time Cycle Clock Cycles Clock CPU Time CPU = ´ = CA: Performance and Power
- 9. 9 CPU Time Example n Computer A: 2GHz clock, 10s CPU time n Designing Computer B n Aim for 6s CPU time n Can do faster clock, but causes 1.2 × clock cycles n How fast must Computer B clock be? 4GHz 6s 10 24 6s 10 20 1.2 Rate Clock 10 20 2GHz 10s Rate Clock Time CPU Cycles Clock 6s Cycles Clock 1.2 Time CPU Cycles Clock Rate Clock 9 9 B 9 A A A A B B B = ´ = ´ ´ = ´ = ´ = ´ = ´ = = CA: Performance and Power
- 10. 10 Instruction Count and CPI n Instruction Count (IC) for a program n Determined by program, ISA and compiler n Average cycles per instruction (CPI) n Depends on program, CPU hardware and compiler n If different instructions have different CPI n Average CPI affected by instruction mix Rate Clock CPI Count n Instructio Time Cycle Clock CPI Count n Instructio Time CPU n Instructio per Cycles Count n Instructio Cycles Clock ´ = ´ ´ = ´ = CA: Performance and Power
- 11. 11 CPI Example n Computer A: Cycle Time = 250ps, CPI = 2.0 n Computer B: Cycle Time = 500ps, CPI = 1.2 n Same ISA n Which is faster, and by how much? 1.2 500ps I 600ps I A Time CPU B Time CPU 600ps I 500ps 1.2 I B Time Cycle B CPI Count n Instructio B Time CPU 500ps I 250ps 2.0 I A Time Cycle A CPI Count n Instructio A Time CPU = ´ ´ = ´ = ´ ´ = ´ ´ = ´ = ´ ´ = ´ ´ = A is faster… …by this much CA: Performance and Power
- 12. 12 How Calculate the 3 Components? n Clock Cycle Time: in specification of computer (Clock Rate in advertisements) n Instruction Count: n Count instructions in loop of small program n Use simulator to count instructions n Hardware counter in spec. register (most CPUs) °CPI: • Calculate: Execution Time / Clock cycle time Instruction Count • Hardware counter in special register (most CPUs) CA: Performance and Power
- 13. 13 CPI in More Detail n If different instruction classes take different numbers of cycles å = ´ = n 1 i i i ) Count n Instructio (CPI Cycles Clock n Weighted average CPI å = ÷ ø ö ç è æ ´ = = n 1 i i i Count n Instructio Count n Instructio CPI Count n Instructio Cycles Clock CPI Relative frequency CA: Performance and Power
- 14. 14 Calculating Average CPI n First read CPIi for each individual instruction (add, sub, and, etc.) n Next use when it’s given or calculate relative frequency fi of each individual instruction n Finally multiply these two for each instruction and add them up to get final CPI 𝐶𝑃𝐼 = % !"# $ 𝑓𝑖 ∗ 𝐶𝑃𝐼𝑖 𝑓𝑖 = 𝐼𝐶𝑖/𝐼𝐶 CA: Performance and Power
- 15. 15 Example with Bonus Points to Earn! Op Freqi CPIi Prod (% Time) ALU 50% 1 .5 (33%) Load 20% 2 .4 (27%) Store 15% 2 .3 (20%) Branch 15% 2 .3 (20%) 1.5 • What if you can make branch instructions twice as fast (CPIbr= 1 cycle) but clock rate (CR) will decrease by 12%? Will it be a speedup or slowdown and how much? Instruction Mix (Where time spent) CA: Performance and Power New CPI = 1.35 Time before change = IC*1.5 / CR = IC*1.5/CR Time after change = IC*1.35 / (0.88*CR) = IC*1.534/CR Time before/Time after = IC*1.5*CR / IC*1.534*CR = 0.978 => 2.2% slowdown Speedup < 1 because the time after is greater than the time before the change
- 16. 16 Another CPI Example n Alternative compiled code sequences using instructions in classes A, B, C Class A B C CPI for class 1 2 3 IC in sequence 1 2 1 2 IC in sequence 2 4 1 1 n Sequence 1: IC = 5 n Clock Cycles = 2×1 + 1×2 + 2×3 = 10 n Avg. CPI = 10/5 = 2.0 n Sequence 2: IC = 6 n Clock Cycles = 4×1 + 1×2 + 1×3 = 9 n Avg. CPI = 9/6 = 1.5 CA: Performance and Power
- 17. 17 CPU Performance Law n Performance depends on n Algorithm: affects IC, possibly CPI n Programming language: affects IC, CPI n Compiler: affects IC, CPI n Instruction set architecture: affects IC, CPI, Tc n Processor microarchitecture (organization): affects CPI, Tc n Technology: affects Tc The BIG Picture Rate Clock CPI Count n Instructio Time Cycle Clock CPI Count n Instructio Time CPU n Instructio per Cycles Count n Instructio Cycles Clock ´ = ´ ´ = ´ = cycle Clock Seconds n Instructio cycles Clock Program ns Instructio Time CPU ´ ´ = CA: Performance and Power
- 18. 18 What Programs Measure for Comparison? n Ideally run typical programs with typical input before purchase, or before even build machine n Called a “workload”; For example: n Engineer uses compiler, spreadsheet n Author uses word processor, drawing program, compression software n In some situations it’s hard to do n Don’t have access to machine to “benchmark” before purchase n Don’t know workload in future CA: Performance and Power
- 19. 19 SPEC CPU Benchmarks n Standard Performance Evaluation Corporation (SPEC) www.spec.org n Elapsed time to execute a selection of programs n Negligible I/O, so focuses on CPU performance n SPEC95: 8 integer (gcc, compress, li, ijpeg, perl, ...) & 10 floating-point (FP) programs (hydro2d, mgrid, applu, turbo3d, ...) n SPEC2000: 11 integer (gcc, bzip2, …), 18 FP (mgrid, swim, ma3d, …) n Separate average for integer and FP n Benchmarks distributed in source code n Compiler, machine designers target benchmarks, so try to change every 3 years CA: Performance and Power
- 20. 20 How Summarize Suite Performance (1/4) n Arithmetic average of execution time of all programs? n But they vary by 4X in speed, so some would be more important than others in arithmetic average n Could add a weights per program, but how pick weight? n Different companies want different weights for their products n SPECRatio: Normalize execution times to reference computer, yielding a ratio proportional to performance = time on reference computer time on computer being rated CA: Performance and Power
- 21. 21 How Summarize Suite Performance (2/4) n If program SPECRatio on Computer A is 1.25 times bigger than Computer B, then B A A B B reference A reference B A e Performanc e Performanc ime ExecutionT ime ExecutionT ime ExecutionT ime ExecutionT ime ExecutionT ime ExecutionT SPECRatio SPECRatio = = = = 25 . 1 • Note that when comparing 2 computers as a ratio, execution times on the reference computer drop out, so choice of reference computer is irrelevant CA: Performance and Power
- 22. 22 How Summarize Suite Performance (3/4) n Since ratios, proper mean is geometric mean (SPECRatio unitless, so arithmetic mean meaningless) n n i i SPECRatio ean GeometricM Õ = = 1 1. Geometric mean of the ratios is the same as the ratio of the geometric means 2. Ratio of geometric means = Geometric mean of performance ratios Þ choice of reference computer is irrelevant! • These two points make geometric mean of ratios attractive to summarize performance CA: Performance and Power
- 23. 23 SPECspeed 2017 Integer Benchmarks on a 1.8 GHz Intel Xeon E5-2650L CA: Performance and Power
- 24. 24 Analysis: CT involves breaking down complex ideas, arguments, or situations into smaller components to understand their structure and relationships. Evaluation: It entails assessing the validity, reliability, and credibility of information, sources, and arguments. This involves considering evidence, biases, and potential shortcomings. Interpretation: CT involves interpreting information and data to derive meaning, identify patterns, and draw conclusions. Inference: It includes making reasonable and well-supported conclusions based on available evidence and information. Problem Solving: CT aids in solving problems by approaching them systematically, considering different perspectives, and evaluating potential solutions. Contextualization: It requires considering the broader context and implications of information, ideas, and decisions. Communication: CT involves effectively conveying ideas, arguments, and analyses to others and engaging in constructive discussions. Skepticism: Critical thinkers approach information with a healthy dose of skepticism, questioning assumptions and seeking evidence. Open-Mindedness: It involves being open to different viewpoints, considering alternative explanations, and being willing to change one's views in light of new evidence. Reflection: It includes reflecting on one's own thinking processes, biases, & assumptions. Critical thinking (CT) involves several key elements (in the words of AI itself): What’s at stake for you here Is a single mean a good predictor of the performance of programs in benchmark suite? How much confidence you can have when using it to compare different processors or to predict future performance on similar apps? It's tough to make predictions, especially about the future. Yogi Berra
- 25. 25 How Summarize Suite Performance (4/4) n Does a single mean well summarize performance of programs in benchmark suite? n Can decide if mean a good predictor by characterizing variability of distribution using standard deviation n Like geometric mean, geometric standard deviation is multiplicative rather than arithmetic n Can simply take the logarithm of SPECRatios, compute the standard mean and standard deviation, and then take the exponent to convert back: n The geometric standard deviation, denoted by σg, is calculated as follows: log σg=[1/n∑n i=1(logxi−logG)2]1/2. n where G=n√x1⋅x2⋅…⋅xn is the geometric mean of SPECRatios (x1 . xn). ( ) ( ) ( ) ( ) i n i i SPECRatio StDev tDev GeometricS SPECRatio n ean GeometricM ln exp ln 1 exp 1 = ÷ ø ö ç è æ ´ = å = CA: Performance and Power
- 26. 26 Example Standard Deviation: (1/3) 0 2000 4000 6000 8000 10000 12000 14000 wupwise swim mgrid applu mesa galgel art equake facerec ammp lucas fma3d sixtrack apsi SPECfpRatio 1372 5362 2712 GM = 2712 GStDev = 1.98 • GM and multiplicative StDev of SPECfp2000 for Itanium 2 Outside 1 StDev Itanium 2 is 2712/100 times as fast as Sun Ultra 5 (GM), & range within 1 Std. Deviation is [13.72, 53.62] CA: Performance and Power
- 27. 27 Example Standard Deviation : (2/3) • GM and multiplicative StDev of SPECfp2000 for AMD Athlon 0 2000 4000 6000 8000 10000 12000 14000 wupwise swim mgrid applu mesa galgel art equake facerec ammp lucas fma3d sixtrack apsi SPECfpRatio 1494 2911 2086 GM = 2086 GStDev = 1.40 Outside 1 StDev Athon is 2086/100 times as fast as Sun Ultra 5 (GM), & range within 1 Std. Deviation is [14.94, 29.11] CA: Performance and Power
- 28. 28 Example Standard Deviation (3/3) - 0.50 1.00 1.50 2.00 2.50 3.00 3.50 4.00 4.50 5.00 wupwise swim mgrid applu mesa galgel art equake facerec ammp lucas fma3d sixtrack apsi Ratio Itanium 2 v. Athlon for SPECfp2000 0.75 2.27 1.30 GM = 1.30 GStDev = 1.74 • GM and StDev Itanium 2 v Athlon Outside 1 StDev Ratio execution times (At/It) = Ratio of SPECratios (It/At) Itanium 2 1.30X Athlon (GM), 1 St.Dev. Range [0.75,2.27] CA: Performance and Power
- 29. 29 Comments on Itanium 2 and Athlon n Standard deviation of 1.98 for Itanium 2 is much higher-- vs. 1.40--so results will differ more widely from the mean, and therefore are likely less predictable for Itanium 2 n Falling within one standard deviation: n 10 of 14 benchmarks (71%) for Itanium 2 n 11 of 14 benchmarks (78%) for Athlon n Thus, the results are quite compatible with a lognormal distribution (expect 68%) n Itanium 2 vs. Athlon St.Dev is 1.74, which is high, so less confidence in claim that Itanium 1.30 times as fast as Athlon n Indeed, Athlon faster on 6 of 14 programs CA: Performance and Power
- 30. 30 Amdahl’s Law ( ) enhanced enhanced enhanced Enh. w/ Enh. w/o overall Speedup Fraction Fraction 1 1 ExTime ExTime Speedup + - = = Best you could ever hope to do: ( ) enhanced maximum Fraction - 1 1 Speedup = ( ) ú û ù ê ë é + - ´ = enhanced enhanced enhanced Enh. w/o Enh. w/ Speedup Fraction Fraction 1 ExTime ExTime F CA: Performance and Power
- 31. 31 Amdahl’s Law Example n New CPU 10X faster n I/O bound server, so 60% time waiting for I/O ( ) ( ) 56 . 1 64 . 0 1 10 0.4 0.4 1 1 Speedup Fraction Fraction 1 1 Speedup enhanced enhanced enhanced overall = = + - = + - = • Apparently, its human nature to be attracted by 10X faster, vs. keeping in perspective its just 1.6X faster CA: Performance and Power
- 32. 32 Question n Speedup = 1 (1 - F) + F S Question: Suppose a program spends 80% of its time in a square root routine. How much must you speed up square root to make the program run 5 times faster? 10 (A) 20 (B) None of the above (C) 100 (D) CA: Performance and Power
- 33. 33 Consequence of Amdahl’s Law n The amount of speedup that can be achieved through parallelism is limited by the non-parallel portion of your program! Speedup Number of Processors Parallel portion Serial portion Time Number of Processors 1 2 3 4 5 CA: Performance and Power
- 34. 34 Parallel Speed-up Examples (1/2) n 10 “scalar” operations (non-parallelizable) n 100 parallelizable operations n Say, element-wise addition of two 10x10 matrices. n 110 operations n 100/110 = .909 Parallelizable, 10/110 = 0.091 Scalar Z1 + Z2 + … + Z10 X1,1 X1,10 X10,1 X10,10 Y1,1 Y1,10 Y10,1 Y10,10 + Non-parallel part Parallel part Partition 10 ways and perform on 10 parallel processing units . . . . . . CA: Performance and Power
- 35. 35 Parallel Speed-up Examples (2/2) n Consider summing 10 scalar variables and two 10 by 10 matrices (matrix sum) on 10 processors Speedup = 1/(.091 + .909/10) = 1/0.1819 = 5.5 n What if there are 100 processors ? Speedup = 1/(.091 + .909/100) = 1/0.10009 = 10.0 n What if the matrices are 100 by 100 (or 10,010 adds in total) on 10 processors? Speedup = 1/(.001 + .999/10) = 1/0.1009 = 9.9 n What if there are 100 processors ? Speedup = 1/(.001 + .999/100) = 1/0.01099 = 91 Speedup w/ E = 1 / [ (1-F) + F/S ] CA: Performance and Power
- 36. 36 Strong and Weak Scaling n To get good speedup on a multiprocessor while keeping the problem size fixed is harder than getting good speedup by increasing the size of the problem n Strong scaling: When speedup is achieved on a parallel processor without increasing the size of the problem n Weak scaling: When speedup is achieved on a parallel processor by increasing the size of the problem proportionally to the increase in the number of processors (Gustafson's law) n Load balancing is another important factor: every processor doing same amount of work n Just 1 unit with twice the load of others cuts speedup almost in half (bottleneck!) CA: Performance and Power
- 37. 37 Other Performance Metrics n MIPS – Million Instructions Per Second • MFLOPS - Million Floating-point Operations Per Second 6 10 ) ( _ ´ = s Time count n Instructio MIPS 6 10 ) ( / _ _ ´ = s Time program ops point Floating MFLOPS • PetaFLOPS - 1015 Floating-point Operations Per Second 15 10 ) ( / _ _ ´ = s Time program ops point Floating PFLOPS CA: Performance and Power
- 38. 38 Pitfall: MIPS as a Performance Metric n MIPS: Millions of Instructions Per Second n Doesn’t account for n Differences in ISAs between computers n Differences in complexity between instructions 6 6 6 10 CPI rate Clock 10 rate Clock CPI count n Instructio count n Instructio 10 time Execution count n Instructio MIPS ´ = ´ ´ = ´ = n CPI varies between programs on a given CPU CA: Performance and Power
- 39. 39 Uniprocessor Performance Constrained by power, instruction-level parallelism, memory latency CA: Performance and Power
- 40. 40 Multiprocessors n Multicore microprocessors n More than one processor per chip n Requires explicitly parallel programming n Compare with instruction level parallelism n Hardware executes multiple instructions at once n Hidden from the programmer n Hard to do n Programming for performance n Load balancing n Optimizing communication and synchronization CA: Performance and Power
- 41. 41 Supercomputing ▪ Today: clusters of multi-core CPUs + GPUs ▪ Frontier: The Exascale-class HPE Cray EX Supercomputer at Oak Ridge National Laboratory (the fastest supercomputer in the world in 2023) ▪ 9,472 AMD Epyc 7453s "Trento" 64 core 2 GHz CPUs (606,208 cores) and 37,888 Instinct MI250X GPUs (8,335,360 cores). ▪ the most efficient supercomputer: 62.68 gigaflops/watt. Space 680 m 2 (7,300 sq ft) Speed 1.194 exaFLOPS (Rmax) / 1.67982 exaFLOPS (Rpeak) Cost US$600 million (est. cost) Purpose Scientific research and development CA: Performance and Power
- 42. 42 Top 5 Supercomputers (TOP500, June 2023) RankSystem Cores Rmax (PFlop/s) Rpeak (PFlop/s) Power (kW) 1 Frontier - HPE Cray EX235a, AMD Optimized 3rd Generation EPYC 64C 2GHz, AMD Instinct MI250X, Slingshot-11, HPE DOE/SC/Oak Ridge National Laboratory United States 8,699,904 1,194.00 1,679.82 22,703 2 Supercomputer Fugaku - Supercomputer Fugaku, A64FX 48C 2.2GHz, Tofu interconnect D, Fujitsu RIKEN Center for Computational Science Japan 7,630,848 442.01 537.21 29,899 3 LUMI - HPE Cray EX235a, AMD Optimized 3rd Generation EPYC 64C 2GHz, AMD Instinct MI250X, Slingshot-11, HPE EuroHPC/CSC Finland 2,220,288 309.10 428.70 6,016 4 Leonardo - BullSequana XH2000, Xeon Platinum 8358 32C 2.6GHz, NVIDIA A100 SXM4 64 GB, Quad-rail NVIDIA HDR100 Infiniband, Atos EuroHPC/CINECA Italy 1,824,768 238.70 304.47 7,404 5 Summit - IBM Power System AC922, IBM POWER9 22C 3.07GHz, NVIDIA Volta GV100, Dual-rail Mellanox EDR Infiniband, IBM DOE/SC/Oak Ridge National Laboratory United States 2,414,592 148.60 200.79 10,096 CA: Performance and Power
- 43. 43 Power Trends n In CMOS IC technology Frequency Voltage load Capacitive Power 2 ´ ´ = ×1000 ×40 5V → 1V n Intel 80386 consumed ~ 2 W n 3.3 GHz Intel Core i7 consumes 130 W n Heat must be dissipated from 1.5 x 1.5 cm chip n This is the limit of what can be cooled by air CA: Performance and Power
- 44. 44 Energy and Power n Dynamic energy n Transistor switch from 0 -> 1 or 1 -> 0 n ½ x Capacitive load x Voltage2 n Dynamic power n ½ x Capacitive load x Voltage2 x Frequency switched n Reducing clock rate reduces power, not energy n Static power consumption n Currentstatic x Voltage n Scales with number of transistors n To reduce: power gating CA: Performance and Power
- 45. 45 Switching Energy: Fundamental Physics Every logic transition dissipates energy. How can we limit switching energy? (1) Reduce # of clock transitions. But we have work to do ... (2) Reduce Vdd. But lowering Vdd limits the clock speed ... (3) Fewer circuits. But more transistors can do more work. (4) Reduce C per node. One reason why we scale processes. V dd 1 2 C V dd E0->1 = 2 V dd 1 2 C V dd E1->0 = 2 C Strong result: Independent of technology. CA: Performance and Power
- 46. 46 0V = Second Factor: Leakage Currents Even when a logic gate isn’t switching, it burns power. Igate: Ideal switches have zero DC current. But modern transistor gates are a few atoms thick, and are not ideal. Isub: Even when this nFet is off, it passes an Ioff leakage current. We can engineer any Ioff we like, but a lower Ioff also results in a lower Ion, and thus a lower maximum clock speed. Intel’s 2006 processor designs, leakage vs switching power A lot of work was done to get a ratio this good ... 50/50 is common. CA: Performance and Power
- 47. 47 Example of Quantifying Power n Suppose 15% reduction in voltage results in a 15% reduction in frequency. What is impact on dynamic power? dynamic dynamic dynamic OldPower OldPower witched FrequencyS Voltage Load Capacitive witched FrequencyS Voltage Load Capacitive Power ´ ´ ´ ´ ´ ´ ´ ´ ´ » = ´ = = 6 . 0 ) 85 (. ) 85 (. 85 . 2 / 1 2 / 1 3 2 2 CA: Performance and Power