FAHBench 2.2

[mpharrigan]

Hi All,

FAHBench 2.2 using the technology of core 21 is available at http://fahbench.github.io/

Feel free to post your scores here! I see the other thread has a curated list as the top post. I'm not going to do that, but if a mod want to volunteer, I can set you up with a page on the website that you can update with a list of scores.

Happy folding

[bruce]

Do we need to keep the old version around for WindowsXP or does FAHBench avoid that problem? (I don't have an XP system to test any more.)

[PS3EdOlkkola]

There are several options available when running FAHBench. What options should we use that have the most relevance?

For example:
* Assuming GPU and OpenCL and Single Precision
- Would you want to see the results from all three work units?
- Should Accuracy Check be enabled?
- For NaN check: Is "disabled" appropriate, or should we choose a certain number of steps instead, and if so, what should they be?
- Is 60 seconds a long enough to runtime to deliver a valid result?

Also, how should the Score be interpreted for each of the runs?

[mpharrigan]

It *may* work on xp if you manually install the vc2015 redistributable (available through microsoft) but I don't have a system to test

Check out http://fahbench.github.io/options.html for a description of the options. If you want a recommendation, just use the defaults

[foldy]

FAHBench 2.2.5 with default settings
Windows 7 64 with AMD driver 16.3.2
AMD HD 7950 @ 880Mhz Score 36.9
AMD R9 280 @ 980Mhz Score 40.5

[Kougar]

Awesome, very glad to see this was updated! Should be much better for cross-comparing to Radeons now

Default Settings
Windows 10 Driver 364.51
NVIDIA Titan Black Stock 65.68

[foldy]

Should be much better for cross-comparing to Radeons now

Why?

[Kougar]

Should be much better for cross-comparing to Radeons now

Why?

The older cores were not well optimized for Radeons, they generally under-performed compared to NVIDIA cards. Core 21 offers much better performance on Radeons by comparison.

With Core 21 projects a Radeon 290X can match a Titan Black in PPD, which is verified by the almost identical 67 score I've seen for someones 290X on a different forum.

[foldy]

I see. For my HD 7950 and R9 280 the score did not change compared to FahBench 1.2

[toTOW]

Previous version of FAHBench (I guess it was 1.2) was based on OpenMM code used in Core 17, and was not affected by performance issues.

This version is based on Core 21 and is performing well on both ATI and NV too. The only differences you may see are from OpenMM optimizations that has been added since then.

[Kougar]

The only difference?? A 290X is able to out-fold a Titan Black, from what I recall they used to only achieve half of that.

[foldy]

I checked FahBench 1.2 numbers and 290x gets single explicit 46 ns/s and Titan Black (780 TI) gets 49 ns/d so they are very similar.
Maybe you recall not the FahBench numbers but the FahCore_18 numbers were AMD performance was bad compared to Nvidia?
But FahBench 1.2 used FahCore_17 and I fixed that AMD issue in FahCore_21. There is no FahBench using FahCore_18

[toTOW]

foldy is right, Kougar is mixing his memories ... Core 17 is not affected by performances issues (so is FahBench 1.2 which use the same code base). Core 18 has terrible performances on ATI GPUs (and we don't have a FahBench that uses it). Core 21 solved the performances issues and we're back to what we had with Core 17 (and the new FahBench uses Core 21 code base)

[foldy]

Maybe this is the right thread to tell the story how a german donor fixed the Core_18 AMD performance issue on Windows for Core_21 using FahBench. I downloaded FahBench 2.0 beta in mid 2015 and run it on my AMD R9 280 but the score was bad compared to FahBench 1.2 matching the FahCore_18 experience.

I wanted to look into this but the FahCores are closed source that means you cannot build it yourself because the source code is not available.
So it is not so easy for a software developer which is not part of the PandeGroup to check what is going wrong.

But FahBench is open source on GitHub so with technical knowledge you can build it from source and get your own FahBench.exe
The common base for both FahCores and FahBench is the OpenMM framework which is also open source.
So I could build FahBench 2.0 beta and OpenMM 6.2 from scratch. And you can even run real FAH work units in FahBench.

What a surprise when my build FahBench 2.0 beta performance score on my AMD R9 280 was good again and the same as FahBench 1.2

So something must be different from my build compared to released FahBench 2.0 beta dlls on GitHub.
We found FahBench 2.0 beta and also FahCore were build at Stanford on a machine having a Nvidia SDK to support OpenCL and CUDA.
However on my machine there was no Nvidia but AMD SDK because I have an AMD R9 280 GPU.
When I build FahBench 2.0 beta/OpenMM 6.2 on my machine using Nvidia SDK I got the bad performance issue again.

So the solution was to build the OpenCL code with AMD SDK. After the developers at Stanford were convinced of this solution a new FahCore_21 beta core was build with AMD SDK for OpenCL and the FahCore_18 introduced AMD bad performance issue was solved.

This shows that donors can also support development if they have the knowledge.
We may have 100000 donors but only 10 developers at Standford.
They may be a little denying if a stranger tries to mess with the experts but if you can prove your facts it will be accepted.
And like donors are familiar with we just give and we don't expect a Thank You

[Kougar]

Ah, got it! Thanks for the correction and the story/background info Foldy! It's always interesting to me to hear the hows and whys behind things