OpenMM 2.0 OpenCL benchmarks

[David Koes]

I did a quick evaluation of OpenMM 2.0/OpenCL/nVidia/ATI and wanted to share the results. I performed a short (100ps) simulation of a small (249 atom) protein with openmm enabled gromacs. I have two high-end video cards, an nVidia GeForce GTX 480 and an ATI Radeon 5970, and an Intel Core i7 3.33GHz 975 CPU. I tried different devices and platforms (using the OPENMM_PLATFORM environment variable). All the latest (non beta) drivers are installed. Here's the results:

Platform Time (s)
nVidia/CUDA 10
nVidia/CUDA/OpenMM1.1 12
nVidia/OpenCL 24
ATI/OpenCL 52
CPU/OpenCL 558

It would be great if other people posted any other benchmarking results to this thread.

[Peter Eastman]

I'm hopeful those OpenCL numbers will get better in the near future. We're working closely with both Nvidia and ATI to understand where the performance problems are coming from. With a combination of OpenMM changes and driver changes, we should be able to match the CUDA numbers.

Peter