FAHBench (OpenMM 5.1)

[tictoc]

Here are a few scores for my 7970

CPU: AMD 1090t @ 3900Mhz
GPU: HD7970 @ 1222Mhz core; 1500Mhz memory
Driver: 13.10_betav2
OS: Windows 7x64

Single Precision:
Explicit -- 41.1148 ns/day
Implicit -- 145.739 ns/day

Double Precision:
Explicit -- 12.1193 ns/day
Implicit -- 22.3136 ns/day

[mdk777]

An OpenMM 5.2 version of Fahbench going to be released ?

[bruce]

An OpenMM 5.2 version of Fahbench going to be released ?

Planned, but no ETA has been suggested.

[bruce]

Well, FYI for anyone interested:

The last few releases of AMD CAT Beta have run the Implicit Solvent test without issue.
My current result with AMD 7970

33.7636 ns/day Explicit
126.75 ns/day Implicit

Unfortunately the FAHClient assigns a mixture of implicit and explicit assignments so core17 will continue to be "advanced" until it's reliable on both.

[PantherX]

Please note that the posts regrading a custom benchmark run using WU Project: 7811 (Run 0, Clone 405, Gen 210) has been moved to its own thread here (viewtopic.php?f=38&t=25229).

EVGA 780 GTX Classified@1201 MHz
Folding@Home Benchmark v.1.2.0 OpenCL-Single

Explicit: 49.678 ns/s
Implicit: 202.33 ns/s

OpenCL-Double

Explicit: 10.0926 ns/s
Implicit: 14.8229 ns/s

I assume that you mean ns/day (which is how I have added it) and not ns/s.

FAHbench on my HD 7950 @ 1000Mhz
Driver 13.4

Explicit 30.1
Implicit error

I added it with the assumption that it is single precision.

Well, FYI for anyone interested:

The last few releases of AMD CAT Beta have run the Implicit Solvent test without issue.
My current result with AMD 7970

33.7636 ns/day Explicit
126.75 ns/day Implicit

These numbers are not unusual. However, testing was run on AMD CAT 13.11 Beta which supports the next generation of cards.

Hopefully this indicates a continuum of FOLDING support.

I have added them with the assumption that they are single precision.

[doesntmatter]

OS -> Windows7 64
GPU -> AMD HD 7950 Boost @stock (1000/1250 - Core/Memory)
Driver -> AMD Catalyst 13.9

(Compared to existing results)
OpenCL Explicit SP: 36.1814 ns/day - like a oced GTX 670
OpenCL Implicit SP: 109.75 ns/day - between HD 7950 and GTX 760

OpenCL Explicit DP: 8.15148 ns/day - between GTX 570 and 580, far away from a HD 7970
OpenCL Implicit DP: 16.166 ns/day - between HD 7950 and HD 7970

So far my intention to use FAHBench as performance check for correct drivers to run FAH did the job.



Question about processing Double Precision, usage of the GPU was dropping to ~64%, clock rate from Boost to base frequency (1000>900), does it slow down the result (which is o.k. in my opinion) ? I think it should be normal, part of calculating. But I wonder nethertheless, thought this Benchmark produces constant load on the GPU, as seen with Single Precision.

[PantherX]

...Question about processing Double Precision, usage of the GPU was dropping to ~64%, clock rate from Boost to base frequency (1000>900), does it slow down the result (which is o.k. in my opinion) ? I think it should be normal, part of calculating. But I wonder nethertheless, thought this Benchmark produces constant load on the GPU, as seen with Single Precision...

A reduction in frequency would slow down the result. Does the GPU Usage drop from ~99% to ~64% or does it stay at 64%? The variability of GPU Usage is a feature of FahCore_17 so it is expected but the drop is usually ~2% so not very noticeable on my GPUs (Nvidia Fermi and Keplers).

[doesntmatter]

Does the GPU Usage drop from ~99% to ~64% or does it stay at 64%?

See detail from the screenshot above, calculation period before 82% of Double Precision:



P.S. Thanks for updating !

[PantherX]

The usage pattern is rather extreme for FahCore_17. However, I personally wouldn't worry about it too much since this occurs only with Double Precision. F@H uses only Single Precision (for both, CPU and GPU) so if they were any issues with Double Precision, it might be a low priority to fix.

[user123]

Gigabyte GT430 back from RMA. Decided to run FAHBench on it
The GT430 was overclocked to 797 MHz GPU, 926 MHz memory. (Default clocks are 730 MHz GPU, 900 MHz memory)
Nvidia driver version WHQL 301.42.



This test took a long time to complete.

[PantherX]

Thanks for your results, I have updated the first post.

BTW, if you think that 1.1186 ns/day took long, image how it felt like with 000.243642 ns/day (AMD Opteron 185)

[bruce]

[The Double Precision] test took a long time to complete.

Double precision performance has never been a priority in consumer products. Business grade GPUs are better suited for DP processing but at a much higher cost.

[Napoleon]

Personally, I look at OpenCL Explicit Single Precision whenever I try to draw any parallels between FahBench performance and folding performance. FYI user123, my Avast antivirus Web Shield reports the image URLs in your posts as malicious. I'm 99.999% sure the following links are in fact OK, but better safe than sorry.

Code: Select all

http://www.imageupload.co.uk/images/2013/12/29/GT430SP.jpg
http://www.imageupload.co.uk/images/2013/12/29/GT430DP.jpg

[PantherX]

Regarding the URL, only WOT has marked it as "Malicious site" while others mostly classify it as Clean Site with a few as Unrated site:
https://www.virustotal.com/en/url/9652e ... 388402346/
https://www.virustotal.com/en/url/ff8a3 ... 388402376/

[Assimilator1]

Well I can't find out if my GTX 260 c216 does support O CL 1.1 or not *shrug*, so seeing as no one's given a benchmark for this oldie I will .

Btw don't you want DP figures as well as SP? A lot of people seem to be posting SP only.

GTX 260 c216 (verify accuracy on, drivers 327.23)

Single Precision

CUDA:
Explicit Solvent:
Implicit Solvent:

Got 'Error loading library D:\FILE STORAGE\Storage\DC\FAHBench_1_2_0/OpenMMCUDA.dll: 126' for above

OpenCL:
Explicit Solvent: 5.58734 ns/day
Implicit Solvent: 31.1255 ns/day

Double Precision

CUDA:
Explicit Solvent:
Implicit Solvent:

Got 'Error loading library D:\FILE STORAGE\Storage\DC\FAHBench_1_2_0/OpenMMCUDA.dll: 126' for above

OpenCL:
Explicit Solvent:
Implicit Solvent:

Got 'Error compiling kernel: ptxas : error : Entry function 'computeNonbonded' uses too much shared data (0x4888 bytes + 0x10 bytes system, 0x4000 max)' for above