-
Notifications
You must be signed in to change notification settings - Fork 3
openmm-hip Testing and Benchmarks #1
Comments
Thank you very much for testing! It's overclocked, right? What is its default frequencies? The specs say that GPU has 1400 and 1750 in boost and memory has 1000. |
Yes those results were with the GPU highly overclocked. At stock, the core clock boosts to 1775-1800MHZ and the memory speed is 1000MHz. Here is a run at stock clocks for comparison. That should fall somewhere right around the expected results on an MI50, since these are all single precision benchmarks. The only change from stock is to set the perf level to high to minimize the noise from the somewhat inconsistent boost algorithm on Vega 20. Average clocks during the below benchmark run were 1770-1790MHz. Full benchmark.py output
OpenCL vs HIP Performance SummarySystemOS: Arch Linux
|
Additional test results with a 6900XT. Improvements are even greater than on the Radeon VII. The 6900XT was tested at default clocks, with only a change in fan speed and setting power to 293W. This did allow the GPU to run at a higher boost clock on the HIP tests, which bump right up against the power limit when it is set at 293W. 1 failed test on test_openmm_hip.sh Test Results
Full benchmark.py output
OpenCL vs HIP Performance SummarySystemOS: Arch Linux
|
Thank you, @tictooc! It's interesting, the hipFFT test fails on RDNA. We added this test because we encountered correctness issues for some FFT sizes on older versions of rocFFT. And now it happens again. I guess we'll need to investigate it further and report to rocFFT developers. |
The same test fails identically on Vega 20 (at least on the Radeon VII) running ROCm 5.1.3. I'll roll back to ROCm 5.0.2, and see if I can find the regression. --Edit-- |
This is not an issue, just wanted to report that the conda version of this plugin along with the StreamHPC/openmm fork are working without issue on ROCm 5.0.2 and the the latest stable kernel(5.16.15).
test_openmm_hip.sh passes all tests.
Test Results
All of the benchmarks except for the amber20-factorix (upstream issue #3391) benchmark run without issue. benchmark.py output using the draft benchmark.py #3386 with a few local changes for HIP and system info
Full benchmark.py output
OpenCL vs HIP Performance Summary
System
OS: Arch Linux
Kernel: 5.16.15
ROCm Version: 5.0.2
OpenMM Version: OpenMM 7.7 | Git Revision: ce22dbef84ec68aa910bbffed0f5e801e76ed9be
CPU: AMD Ryzen Threadripper 3960X @ 4.2GHz
GPU: AMD Radeon VII @ 2120core|1200mem
The text was updated successfully, but these errors were encountered: