SPEC just released a new version of the Chauffeur WDK, a development kit that makes it easy to write your own benchmarks to measure power and performance. Why would anyone want to write their own benchmark? SPEC and other organizations already offer benchmarks for energy efficiency, and these are certainly a good starting point. But your workload matters. How much?
The SPEC ACCEL™benchmark is a suite of computationally intensive parallel applications executed to measure performance of the accelerator, host CPU, memory transfer between host and accelerator, support libraries and drivers, and compilers. 15 applications are executed 3 times each, with moderate tuning and with aggressive tuning to yield a base and peak score. This benchmark should be increasingly important as the industry embraces hybrid computing for applications like machine learning. For my purposes today the key point is that different workloads are executed, and that the benchmark allows energy consumption to be measured using the SPEC Power methodology.
I examined one result, for an ASUS P9X79 Motherboard published March 2014. It’s a single chip Intel Core i7-3930K running at 3.2-3.8 GHz, with an NVIDIA Tesla K40c accelerator. It was tested using OpenACC with the PGI Accelerator Server Complete, Release 14.2 compiler.
SPECaccel_acc_base =2.21 SPECaccel_acc_energy_base =2.88 SPECaccel_acc_peak =2.32 SPECaccel_acc_energy_peak =2.98
Each of three runs yields an average power consumed over the course of that run. Here are the median watts used for 3 of the 15 benchmarks.
Benchmark Base Peak 314.omriq 320 320 351.palm 217 225 363.swim 243 265
351.palm, a large-eddy simulation model for atmospheric and oceanic flows, consumed 217 watts with base tuning. By contrast 314.omriq, which computes MRI image reconstruction, consumed 320 watts. That’s 47% higher power consumption for the medical imaging workload. Even the same workload can differ in power consumption. 363.swim is a finite-difference approximation of the shallow-water equations for weather prediction. Compiled with peak tuning it uses 9% more power than when running with base tuning. However it finishes the problem faster with peak tuning, so the total amount of energy consumed, 18.1 k-Joules, is 25% lower than the energy consumed to solve the problem with base tuning, 24.2 k-Joules.
Your workload matters. How much energy and how many dollars are your computing solutions costing you?