We have been aggressively working on OpenMM (the key code used in the FAH GPU cores), creating new algorithms to increase performance on NVIDIA and AMD GPUs. The results have been pretty exciting. With OpenMM 5.1 (vs OpenMM 5.0, used in the current core 17 release), we are getting about a 2x speed up on typical FAH WU calculations, which will lead to an automatic 2x increase in PPD once this software is out of beta testing and integrated into core 17.
There's a lot of testing to do and it's very possible that these numbers will change, but the results were so exciting that I wanted to give donors a heads up. Here's some numbers that we're seeing:
OpenCL running on the GTX 680: The first 2 columns are nanoseconds per day (i.e. how much science gets done in a GPU day) and the 3rd column is the speedup of 5.1 over 5.0.
|
Type of Calculation |
OpenMM 5.0 |
OpenMM 5.1 |
Speedup |
|
Implicit hbonds |
92 |
134 |
1.46 |
|
Implicit hangles |
153 |
209 |
1.36 |
|
RF hbonds |
31.4 |
78.1 |
2.49 |
|
RF hangles |
58 |
113.0 |
1.95 |
|
PME hbonds |
19.6 |
41.5 |
2.12 |
|
PME hangles |
37.3 |
66.9 |
1.79 |
OpenCL running on a Radeon HD 7970: The first 2 columns are nanoseconds per day (i.e. how much science gets done in a GPU day) and the 3rd column is the speedup of 5.1 over 5.0.
|
Type of Calculation |
OpenMM 5.0 |
OpenMM 5.1 |
Speedup |
|
Implicit hbonds |
87 |
120 |
1.38 |
|
Implicit hangles |
96 |
104 |
1.09 |
|
RF hbonds |
33.5 |
83.5 |
2.49 |
|
RF hangles |
51.8 |
90.2 |
1.74 |
|
PME hbonds |
21.8 |
49.3 |
2.26 |
|
PME hangles |
34.6 |
63.0 |
1.82 |
Note that "PME hbonds" is likely the most common calculation that we plan to run in the near term with core 17. We're very excited about the way this is shaping up and think that donors would be curious to know where this is going.