![Page 1: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/1.jpg)
Power Measurements for Computer and Computer Architecture Research
Hyesoon Kim
![Page 2: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/2.jpg)
Power Measurements
CPUGPUHDDothers
2
0
20
40
60
80
100
120
140
160
App A App B App C App D
![Page 3: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/3.jpg)
Measurements
• Power measurements – System Input power– Module power – Board level
• Temperature measurements
3
![Page 4: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/4.jpg)
System Power Measurement
4
Power = Dynamic power + static power Dynamic power ∝ dynamic activity Static power ∝ Area, Temperature
![Page 5: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/5.jpg)
Module Power: Memory
5
![Page 7: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/7.jpg)
Module Power: CPU
7
![Page 8: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/8.jpg)
Board Level Power Measurements
• Probe Station
8
Board level Multi-point IR, RFmeasurements
![Page 9: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/9.jpg)
CASE STUDY: GPU POWER MODEL [ISCA’10]
9
![Page 10: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/10.jpg)
Overview of GPU Architecture
10
Software-managed cache
SIMD Execution Unit inside SM
Hong, Kim, ISCA’10
![Page 11: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/11.jpg)
Power Model
11
Power contributions: SMs, Memory and Idle-power
Streaming multiprocessor (SM) decomposition
Memory: Global memory systemHong, Kim, ISCA’10
![Page 12: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/12.jpg)
Training Benchmarks To Find Parameters
12
Training benchmarks stress particular units of processors, memory
Hong, Kim, ISCA’10
![Page 13: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/13.jpg)
Power Model
13
Power parameters used
Units inside SM are per SM granularity
Initial increase from idle to low accessrate causes large power consumption for some units
Hong, Kim, ISCA’10
![Page 14: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/14.jpg)
Predictions
14Hong, Kim, ISCA’10
![Page 15: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/15.jpg)
Power Consumption Breakdown
15Hong, Kim, ISCA’10
![Page 16: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/16.jpg)
Usage 1: Find Optimal Number of Cores
16
Performance, Power and Efficiency vs. Number of Active Cores
Type 1 utilizes all the cores (Performance increases linearly)
Performance for type 2 saturates (Due to bandwidth limitations)
Hong, Kim, ISCA’10
![Page 17: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/17.jpg)
Results
17
Energy savings using optimal number of cores based on IPP system (NVIDIA GTX 280 and power gating GPUs)
![Page 18: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/18.jpg)
Usage 2: Dynamic Mapping
• Distribute work between CPU and GPU for power and performance
GPU CPU
CASE 1
CASE 2Luk & Hong & Kim MICRO’09
![Page 19: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/19.jpg)
Energy Consumption
19Luk & Hong & Kim MICRO’09
![Page 20: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/20.jpg)
TEMPERATURE MEASUREMENTS
20
![Page 21: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/21.jpg)
Using Thermal Camera
21http://users.soe.ucsc.edu/~renau/rpapers.html
![Page 22: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/22.jpg)
Using Thermocouples
22
With Thermocouples, the original case can be assembled so we can measure power & temp behavior with the original heat sink
![Page 23: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/23.jpg)
On-going Researches Topcis
• Thread scheduling policies • DVFS managements
H: running hot jobs, C: cold jobs or low voltage, turn-off power• Multi-core lay out design to reduce thermal gradient, effective
thermal transfers
• Leakage power model (combine with themperature model) 23
H C C HC C C CC C C CH C C H
H H C CH H C CC C C CC C C C
Figures are from http://www.ece.ubc.ca/~aamodt/papers/gyuan.micro2009.pdf
![Page 24: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/24.jpg)
Conclusions
• Power & temperature measurements opens an new architecture research – Architecture design (not only performance but
also power) – Application scheduling, resource managements – Dynamic compilation system
• Experimental measurements and simulations together can extend more wider range of research
24
![Page 25: Power Measurements for Computer and Computer Architecture](https://reader031.vdocuments.net/reader031/viewer/2022021209/620636078c2f7b1730057785/html5/thumbnails/25.jpg)
Dream Lab
• Power measurement systems – Current –probed oscilloscopes– IR measurement systems– Current measurement controllers
• Probe stations • Temperature measurements
– Thermal image camera, oil cooling system– Multiple of thermocouple measurement
systems
• Facility– Humidity, temperature controlled room – High current power supply
• Custom boards, custom riser cards
25