Power of Realtime 3D-Rendering. Raja Koduri

Transcription

1 Power of Realtime 3D-Rendering Raja Koduri 1

2 We ate our GPU cake - vuoi la botte piena e la moglie ubriaca And had more too! 16+ years of (sugar) high! In every GPU generation More performance and performance-per-watt More programmability, precision and features While maintaining compatibility with 8+ generations of APIs 2

3 Basic Equations Chip Power = Static Power + Dynamic Power System Power = CPU + GPU + Other Static power is leakage of inactive transistors Static Power = N*V*e -Vt Dynamic power is from active switching transistors Dynamic Power = A*N*C*F*V 2 N - Number of transistors V - Voltage Vt - Thresh-hold Voltage A - Activity Factor F - Frequency C - Capacitance per transistor Energy = Power(Watts) * Time 3

4 System Categories and GPU Power Limits 0-10 Watts TDP Watts TDP Watts TDP Desktops Mobile Devices (Phones, tablets etc) Mobile Computers (Computers with a battery) Desktops, Servers, Consoles etc (Always plugged in) TDP - Thermal design power Maximum amount of power that the thermal system can sustain 4

5 Disruptive transition - Chip technology limits Sub Moore s law scaling of performance-per-watt N wants to go up, but Sub-Moore scaling on C and V. GPUs scale better than CPUs Beware of the Fine Print though 5

6 Disruptive transition - Market shift Towards lower power computers Battery life Thermal limits Acoustics $-per-kilowatthour Green 6

7 System categories & Volumes 300 W Desktops 50 W Mobile Computers 10 W 0 W Mobile Devices HW vendors compelled to succeed in mobile markets 7

8 CPU v/s GPU v/s FixedFunction Chip Power = N*V*e -Vt + A*N*C*F*V 2 CPUs Prioritize Frequency - higher V and Vt Spend N for caches, cores, flexibility and compute quality GPUs Lower Frequency and Voltage Spend N for shaders, textures, pixels etc FixedFunction Lowest N, F and V for a given task 8

9 CPU v/s GPU If your workload parallelizes well on a GPU Use the GPU It s only a win if GPU_Workload_power*GPU_Executiontime + CPU_GPU_feeding_power*CPU_GPU_feeding_time is less than CPU_Workload_power*CPU_Executiontime Optimize for System Energy 9

10 A Peek inside the power management black box 10

11 State of Art - Static Power Management Static Power = N*V*e -Vt Power off unused areas(power gating) When not in use Static Power ~= 0 Fine print Latency with power toggle (few nano-seconds to a few hundred microseconds) Too aggressive switching may cause performance problems, too conservative switching may lead to wasted power 11

12 Dynamic Power Management Dynamic Power = A*N*C*F*V 2 Primary OS+HW strategy is to control F & V Based on history of A Applications have control of Activity(A) Lower A, leads to lower F and V and lower power. Fine Print Switching F&V states can range between a few milliseconds to few seconds! 12

13 Activity->Performance->Power 1.0 Activity/Performance/Power sample illustration Activity/Frequency/Power State Switch latency 0 Time Activity Frequency Power 13

14 Power Management system diagram Applications GPU API/Runtime GPU UserMode Drivers GPU Power Management Driver GPU Kernel Mode Driver Drivers Power u-controller GPU 14

15 Basic Application Optimizations for Power Tip 1 Control frame rate to minimum desired Pumping out more frames than user can see is wasteful anyways 15

16 Wasteful frame-rate - case study Frames-per-sec Power in Watts 50 Watt System 300 Simple rendering (like menus) Complex rendering (like game play) 100 Power

17 Case study - What happens in a 25 watt system? Frames-per-sec Power in Watts 25 Watt System 300 Simple rendering (like menus) fps app Complex rendering (like game play) 100 Thermal limit Power

18 Real world good example 18

19 Basic Application Optimizations for Power Tip 2 Optimize the frame rendering time to a minimum Don t stop optimizing when you hit your minimum frame rate targets! 19

20 Tip 2 - Ideal rendering algorithm case Case1 Energy 4*Pmax+12*Pstatic Pstatic is near zero in power optimized GPUs, So 60 Case1 Energy 4*Pmax Activity Time in Milliseconds 20

21 Tip 2 - Sub-optimal rendering algorithm case Case1 Energy 4*Pmax Case2 Energy 16*Pmax 60 Activity Time in Milliseconds 21

22 Basic Application Optimizations for Power Tip 3 Don t scatter work in a frame (coalesce) Insufficient idle intervals for power-state reduction 22

23 Don t scatter work in a frame Activity Time in Milliseconds May not be enough idle time for switching to lower power 23

24 Basic Application Optimizations for Power Tip 4 Avoid spin-loops Eg:- CPU waiting on GPU This looks like real work to CPU 24

25 Advanced topics and Research Opportunities 25

26 Scheduling for power optimization A complex subject Dynamic scheduling systems based on power and thermal feedback Hardware v/s Software schedulers Scheduling CPU and GPU Many more topics 26

27 FixedFunction Revenge Premature declaration of death of fixed-function in GPU hardware What new candidates can we move to FixedFunction? What interface principles should FixedFunction hardware adopt to be mainstream programmer friendly? 27

28 Advanced power and thermal management models Can we build Predictive models to augment current reactive models? Should there be APIs for apps to influence power states and monitor feedback? 28

29 Plenty of opportunity for new power optimized rendering approaches 2560x1600 ~ 250 MPixels/Sec A 25W GPU today 5800 MPixels/Sec MTexels/Sec MFLOPS 29

30 Backup 30

31 Static Power Management Static Power = N*V*e -Vt Reduction Choices Reduce N, but that reduces capabilities Reduce V, but limits performance and not in your control(process limits) Reduce Vt, Slower transistor, lower performance Dominant battery life factor for common usage scenarios 31

32 Tip 2 - Anomalous Case Case1 Energy 4*Pmax 80 Case3 Energy 8*0.7*P? Activity Time in Milliseconds 32