How a single chip causes massive power bills GPUSimPow: A GPGPU power simulator

Abstract

Modern GPUs are true power houses in every meaning of the word: While they offer general-purpose (GPGPU) compute performance an order of magnitude higher than that of conventional CPUs, they have also been rapidly approaching the infamous "power wall", as a single chip sometimes consumes more than 300W. Thus, the design space of GPGPU microarchitecture has been extended by another dimension: power. While GPU researchers have previously relied on cycle-accurate simulators for estimating performance during design cycles, there are no simulation tools that include power as well. To mitigate this issue, we introduce the GPUSimPow power estimation framework for GPGPUs consisting of both analytical and empirical models for regular and irregular hardware components. To validate this framework, we build a custom measurement setup to obtain power numbers from real graphics cards. An evaluation on a set of well-known benchmarks reveals an average relative error of 11.7% between simulated and hardware power for GT240 and an average relative error of 10.8% for GTX580. The simulator has been made available to the public [1].

ALUPower: Data Dependent Power Consumption in GPUs

“…By using a Fermi architecture card the test code could be written using asfermi. To measure the energy consumption, a power measurement test-bed similar to the ones used in GPUSimPow [6] and GPUWattch [5] was used. The NVidia CUDA command line profiler was used to gather kernel start and end times.…”

“…GPUWattch [5] and GPUSimPow [6] are power simulators for GPUs. Both simulators are based on gpgpu-sim and extend it with a McPAT-based GPU power model.…”

“…Even for the FADD instruction our energy model is 60.5% more accurate. The constant energy model has different coefficients for different instruction types and odd/even warps, while GPUWattch and gpusimpow [5], [6] only differentiate between integer and floating point instruction. These even simpler models would perform even worse than the constant energy model employed here.…”

“…Only the data values processed are different. The activity factors used by simulators such as GPUWattch [5] or GPUSimPow [6] for all three kernel executions would be completely identical. As a result the predicted power consumption would also be identical.…”

GPGPU workload characteristics and performance analysis

“…In contrast to this, we study power consumption at component level. The recent release of GPUSimPow [6] and GPUWattch [7], GPUs power estimation tools has enabled in-depth exploration of GPUs power consumption. Using metrics to understand workload characteristics is not new [8], [9], [10], [11], but none of them study the correlation between workload metrics and components power.…”