ARM Intelligent Power Allocation

Transcription

1 ARM Intelligent Power Allocation 1

2 Agenda Background and Motivation What is ARM Intelligent Power Allocation? Results Status and Conclusions 2

3 Power Consumption Scenarios The illustration to the right depicts 3 distinct types of power consumption: 1. Short term responsiveness Power consumption allowed to significantly exceed sustainable max power of the SoC for relatively short periods E.g. Web Browsing, UI update, HDR processing (first few shots) Power/Performance Burst for responsiveness T >= Tjmax, Tskin Sustained performance Max Sustainable Power Power Optimised Low End 2. Sustained workload Goal is to maximise performance at, or within, the sustainable power envelope Adapt performance allocation to major power consuming entities E.g. Gaming, long benchmarks, video processing Energy constrained use cases Operates under the radar of intelligent power allocation E.g. low power audio A successful thermal management scheme will maximize the user experience for types1 and 2. Time 3

4 Power Consumption Scenarios The illustration to the right depicts 3 distinct types of power consumption: 1. Short term responsiveness Power consumption allowed to significantly exceed sustainable max power of the SoC for relatively short periods E.g. Web Browsing, UI update, HDR processing (first few shots) Power/Performance Burst for responsiveness T >= Tjmax, Tskin Sustained performance Max Sustainable Power Power Optimised Low End 2. Sustained workload Goal is to maximise performance at, or within, the sustainable power envelope Adapt performance allocation to major power consuming entities E.g. Gaming, long benchmarks, video processing Energy constrained use cases Operates under the radar of intelligent power allocation E.g. low power audio A successful thermal management scheme will maximize the user experience for types1 and 2. Time 4

5 Power Consumption Scenarios The illustration to the right depicts 3 distinct types of power consumption: 1. Short term responsiveness Power consumption allowed to significantly exceed sustainable max power of the SoC for relatively short periods E.g. Web Browsing, UI update, HDR processing (first few shots) Power/Performance Burst for responsiveness T >= Tjmax, Tskin Sustained performance Max Sustainable Power Power Optimised Low End 2. Sustained workload Goal is to maximise performance at, or within, the sustainable power envelope Adapt performance allocation to major power consuming entities E.g. Gaming, long benchmarks, video processing Energy constrained use cases Operates under the radar of intelligent power allocation E.g. low power audio A successful thermal management scheme will maximize the user experience for types1 and 2. Time 5

6 Power Consumption Scenarios The illustration to the right depicts 3 distinct types of power consumption: 1. Short term responsiveness Power consumption allowed to significantly exceed sustainable max power of the SoC for relatively short periods E.g. Web Browsing, UI update, HDR processing (first few shots) Power/Performance Burst for responsiveness T >= Tjmax, Tskin Sustained performance Max Sustainable Power Power Optimised Low End 2. Sustained workload Goal is to maximise performance at, or within, the sustainable power envelope Adapt performance allocation to major power consuming entities E.g. Gaming, long benchmarks, video processing Energy constrained use cases Operates under the radar of intelligent power allocation E.g. low power audio A successful thermal management scheme will maximize the user experience for types1 and 2. Time 6

7 Power Consumption Scenarios The illustration to the right depicts 3 distinct types of power consumption: 1. Short term responsiveness Power consumption allowed to significantly exceed sustainable max power of the SoC for relatively short periods E.g. Web Browsing, UI update, HDR processing (first few shots) Power/Performance Burst for responsiveness T >= Tjmax, Tskin Sustained performance Max Sustainable Power Power Optimised Low End 2. Sustained workload Goal is to maximise performance at, or within, the sustainable power envelope Adapt performance allocation to major power consuming entities E.g. Gaming, long benchmarks, video processing Energy constrained use cases Operates under the radar of intelligent power allocation E.g. low power audio A successful thermal management scheme will maximize the user experience for types 1 and 2. Time 7

8 Current Linux Thermal Framework Thermal trip mechanism Trip2 Trip1 Designed to turn on fans, and/or reduce operating frequency Activates on temperature overshoot Does not accommodate power partitioning between different parts of SoC: GPU, CPU, DSP, video Reactive and static 8

9 IPA Thermal Management technology Proactive vs. Reactive Thermal Management Fixed response to temperature crossing a threshold (Reactive) vs. continuously adapting response based on power consumption and thermal headroom (Proactive) Static vs. Dynamic Partitioning A fixed policy for each component based on prior knowledge of its power consumption vs. a dynamic policy that takes into account component behaviour over time Maximising performance within thermal limits is critical Closed-loop control to dynamically allocate power budget, taking into account: Thermal headroom Real-time performance requirements 9

10 Agenda Background and Motivation What is ARM Intelligent Power Allocation? Results Status and Conclusions 10

11 ARM Intelligent Power Allocation Power to Heat Tdie Tskin Performance Requests big LITTLE GPU Real-time CPU & GPU Performance requests SoC SoC IPA Elements: Temperature control Power estimation Performance allocation big LITTLE GPU Allocated Performance Dynamic Allocation by: Performance required Thermal headroom 11

12 Thermal Management Approach Granted Performance Power Arbiter performs two tasks: 1. Keeps system within thermal envelope Proactive power budget via closed loop control Exploits thermal headroom CPU GPU Perf request Perf metrics Granted Performance Perf request Power model Power model Delta Power Estimate Power Estimate Delta Power Estimate PID Power Arbiter Policy Config 2. Dynamic power allocation per actor, based on Performance demand Power models Perf metrics Granted Performance Perf request Power model Power Estimate Delta Power Estimate Power arbitration Temp. Monitor 12 Actors Perf metrics Power Estimate

13 PID Controller and Power Allocation K_d control_temp e t x K_i D tdp Performance Requests Actor 1 Actor N S e e x I S S Pmax Power Allocation Actor1..N weight Actor1..N maxpower x P -rest_of_soc_power Actor 1 Actor N Performance Granted -Temp K_po/pu PID Controller Provides a power budget, based on current and control temp 13

14 PID Controller and Power Allocation K_d control_temp S e -Temp x K_po/pu e t K_i Power Allocation based on: Requested performance e x S Power budget I Policy Produces final output power x D P tdp S Pmax -rest_of_soc_power Performance Requests Actor 1 Actor N Power Allocation Actor 1 Actor N Performance Granted Actor1..N weight Actor1..N maxpower 14

15 Agenda Background and Motivation What is ARM Intelligent Power Allocation? Results Status and Conclusions 17

16 Configuration Platform: Odroid XU3: 4xA15, 4xA7 ARM Mali-T628 Aims: Target control temperature 65 o C Critical temperature 69 o C Static policy: GPU: 4 Trip points, one for each frequency At each trip point, max frequency is reduced Trips at 56, 57, 58, 59 o C A15/A7: 3 trip points at 60, 62, 64 o C IPA: Control temp = 65 o C Hotplug actioned if reached 69 o C Higher priority for A7 18

17 Intelligent Power Allocation on Games Unity The Chase: T/ C 31% higher FPS with IPA Time Enlighten Transporter: T/ C 8% higher FPS with IPA Time 19

18 Intelligent Power Allocation SOLID Temperature Control Unity the Chase: T/ C Enlighten Transporter: T/ C 20

19 % Improvement on Consecutive Benchmark Runs IPA vs. Static 21

20 Running Frequency Intelligent Power Allocation in Action IPA dynamically allocates power to CPU clusters or GPU, based on load Temperature determines available power You set Temperature, IPA caps Frequencies Load determines how power is divided between CPUs (big and LITTLE) and GPU GLB Trex HD [3 Runs] Max OPP big Max OPP LITTLE big LITTLE GPU Max OPP GPU Median filtered chart for clarity Time (s) 22

21 Running Frequency Intelligent Power Allocation in Action IPA dynamically allocates power to CPU clusters or GPU, based on load Temperature determines available power You set Temperature, IPA caps Frequencies Load determines how power is divided between CPUs (big and LITTLE) and GPU GLB Trex HD [3 Runs] Max OPP big Max OPP LITTLE big LITTLE GPU Max OPP GPU 23 Median filtered chart for clarity Device is still cool There are no constraints on frequency Every actor runs at max frequency Time (s)

22 Running Frequency Intelligent Power Allocation in Action IPA dynamically allocates power to CPU clusters or GPU, based on load Temperature determines available power You set Temperature, IPA caps Frequencies Load determines how power is divided between CPUs (big and LITTLE) and GPU GLB Trex HD [3 Runs] Max OPP big Max OPP LITTLE big LITTLE GPU Max OPP GPU Median filtered chart for clarity High load on GPU Low load on CPU GPU gets most of the power Time (s) 24

23 Running Frequency Intelligent Power Allocation in Action IPA dynamically allocates power to CPU clusters or GPU, based on load Temperature determines available power You set Temperature, IPA caps Frequencies Load determines how power is divided between CPUs (big and LITTLE) and GPU GLB Trex HD [3 Runs] Max OPP big Max OPP LITTLE big LITTLE GPU Max OPP GPU 25 Median filtered chart for clarity High load on CPU Low load on GPU CPU gets most of the power Time (s)

24 Running Frequency Intelligent Power Allocation in Action IPA dynamically allocates power to CPU clusters or GPU, based on load Temperature determines available power You set Temperature, IPA caps Frequencies Load determines how power is divided between CPUs (big and LITTLE) and GPU GLB Trex HD [3 Runs] Max OPP big big LITTLE GPU Max OPP LITTLE Max OPP GPU 26 Median filtered chart for clarity As the device gets hotter IPA reduces available power to actors to maintain temperature control Time (s)

25 Agenda Background and Motivation What is ARM Intelligent Power Allocation? Results Status and Conclusions 27

26 IPA upstreaming Status ARM has produced a Linux OS implementation of IPA Demonstrates key concepts and shows benefit on real workloads ARM contributing IPA to Linux kernel RFCs have been sent to linux-pm Latest: v5 RFC: Supported by Mali DDK Check out IPA code for ARM Juno release (includes GPU integration) (tag lsk-3.10-armlt-juno ) Upstream timescales are subject to variation, depending on community feedback ARM, Linaro, ARM partners all contributing & reviewing 28

27 Conclusions ARM Intelligent Power Allocation is designed to maximise performance in the thermal envelope: Proactively adjusts available power budget, based on device temperature Allocates power dynamically among actors, based on performance demand ARM Intelligent Power Allocation shows solid thermal control and better performance than existing Linux thermal framework We are providing Intelligent Power Allocation to the community to as baseline for best-in-class thermal management 29

28 Thank You The trademarks featured in this presentation are registered and/or unregistered trademarks of ARM Limited (or its subsidiaries) in the EU and/or elsewhere. All rights reserved. Any other marks featured may be trademarks of their respective owners 30