We’ve been pushing hard to improve the performance of OpenMM, especially in OpenCL as it’s now used in Folding@home. We’ve got some great news hot off of the presses. These are the benchmarks described at http://wiki.simtk.org/openmm/BenchmarkOpenMMDHFR. They’re using the very latest OpenMM code, what will be in OpenMM 6.3. They’re using CUDA 6.5 and running on Titan X. All numbers are in ns/day.
| Benchmark Calculation | CUDA | OpenCL |
| Implicit, 2 fs | 471 | 366 |
| Implicit, 5 fs | 684 | 589 |
| Explicit-RF, 2 fs | 305 | 265 |
| Explicit-RF, 5 fs | 508 | 460 |
| Explicit-PME, 2 fs | 161 | 164 |
| Explicit-PME, 5 fs | 318 | 354 |
We’re especially pleased with those OpenCL PME numbers. OpenMM Lead Developer Peter Eastman has put a lot of work into that for this release, and it now is actually faster than CUDA (For the Titan X). Curiously, that is not the case on GTX 980. It’s still slower than CUDA there, although it comes a lot closer than it used to.
This will be spun into an updated Folding@home core. The upshot for GPU donors is that PPD for that new core should increase, due to the expanded capabilities of the new code.
It’s important to stress that SMP/CPU donors aren’t left out of new performance (and therefore PPD) updates either: FAH Lead Developer Joseph Coffland has been working hard on a new Gromacs core and that should also see performance benefits, as we roll out AVX support for FAH.