Folding Forum

Open source is awesome! Note that fahbench includes an (almost) verbatim copy of the core code with the secret checksumming security stuff stripped out and a gui laid on top.

To complete the history above: we not build different cores for nvidia and amd against their respective opencl libraries. It would be nice if we could just use amd app sdk, but that caused problems with older nvidia cards or drivers (can't remember which)

I do remember : minimum OpenCL requirements. All actual FAH software requires OpenCL 1.1, but first alternative builds required OpenCL 1.2 which broke compatibility on many very capable cards ... and since OpenMM code is written with 1.1 in mind, I think that going to 1.2 or 2.0 didn't improve performance or functionality.

FAHBench 2.2.5 with default settings
Windows 7 64 with NVIDIA driver 364.51

Geforce GTX980 @ 1418MHz / 4000MHz
Score dhfr 76.0373
Score dhfr-implicit 323.213
Score nav 11.4755

Just installed 2.2.5 and ran stock settings on my GTX 1080 MSI Gaming X. Score = 124.693

FAHBench 2.2.5 with default settings
Windows 10 LTSB-N x86-64 with NVIDIA driver 368.81

GeForce GTX 780 Ti K|NGP|N @ 1300mV, 1409MHz
Score dhfr 72.2956

Passed 24 hours Folding @ 1200mV, 1333MHz / 7200MT/s

FAHBench 2.2.5 with default settings
Windows 10 Home x64
Nvidia GTX 1060 6GB, OpenCL, single precision

WU dhfr, Score 79.5192
WU dhfr-implicit, Score 351.163, Scaled Score 28.8185
WU nav, Score 9.82658, Scaled Score 86.5147

I am putting together a price/performance comparison of running machine personally compared to cloud services. I am using the FAH Benchmark (http://fahbench.github.io/) to get performance information from the systems. This reports performance as number of simulated nanoseconds per day run on that particular system for a particular configuration. The configuration parameters for the benchmark are precision (single or double) and if the solvent e.g. water is explicitly or implicitly modeled. There are large performance differences between the configurations on different cards. These differences actually change the results of the analysis depending on what real projects use.

Which brings me to the question: Do we know what configuration of precision and solvent that real work units use?

As far as I remember, all simulations use explicit solvent. Most of them only require single precision, but a little number use "mixed precision" (which means that most of the computations are single precision, but a few operations require double).

In the FahBench settings the "nav" work unit is a real work unit.

Since it is possible to add your own data to FAHBench, I imported some real work units from FAH

Thank you for the replies! Any idea what be the correct sub-forum to post the analysis on?

Could you point me at the documentation on how to import work units from FAH? Or is it fairly simple?

Here's how to add our own WUs to FAHBench :
- in \share\fahbench\workunits create a folder for your test WU
- in this folder, put the following files integrator.xml, state.xml and system.xml copying them from the core WU data files (\work\01\01).
- in this folder, create a wu.json file following the example of other sets

If everything is setup fine, your addition should appear on next FAHBench restart.