FAHBench (OpenMM 5.1)

[bruce]

Double precision in FAH is pointless?

FAH is specifically designed to run on consumer-grade hardware rather than requiring business-grade hardware. Since many consumer-grade GPU have no Double Precision or they're have extremely limited capability, it doesn't make sense to require DP. In almost all cases, it adds no benefit. In those rare cases where it might provide a small benefit, Development can generally spend a little extra effort programming around the limitations that happen to be imposed by avoiding the use of DP. In other words, when the final versions of FAH's code are released, DP is pointless -- and, in fact, using DP would slow down the analysis, even if they had used it in the lab to test upcoming code.

It's a very considered decision.

[Ascetus]

ok, this my RTX2060
OpenCL SP
Explicit:
113.341 ns/day
Implicit:
603.647 ns/day

[bruce]

FAH now uses "mixed precision" (almost) all of the time. That's different than running exclusively in SP or exclusively in DP andd isn't overtly supported by FAHBench. Assuming that 5% of the calculations are DP and 95% are SP (i don't know the percentages), the only real conclusion is that full DP is much slower but some portions of the calculations must be done using those slower operations. The effective speed would be slightly slower than the reported values but with the increased accuracy approximating full DoublePrecision.

[foldy]

FAHbench v1.2.0 uses old OpenMM 5.1 which is not used anymore in FAH GPU cores.

Use FAHbench v2.3.2 using OpenMM 6.3 which matches current FAHcore_21
https://github.com/fahbench/fahbench/releases

[Ascetus]

FAHbench v1.2.0 uses old OpenMM 5.1 which is not used anymore in FAH GPU cores.

Use FAHbench v2.3.2 using OpenMM 6.3 which matches current FAHcore_21
https://github.com/fahbench/fahbench/releases

I know, but the old-version benchmark still clearly shows the difference between the various devices. I don’t really like the new benchmark. It is less generalized. More specialized and "bloated" in the settings.

[bruce]

There have been a number of important changes when OpenMM was upgraded. If you prefer the older (unbloated) GUI version of the FAHBench code, I suggest that you Investigate more fully and report back. All of the software involved is OpenSource. Please download OpenMM 6.3 (or greater) and FAHBench 1.2.0. Link the two to create your own custom version of FAHBench. That will allow you the best of both worlds.

Then you can submit the results of your suggested changes to the developers of FAHBench.

[Assimilator1]

Is their not an updated version of this thread?
I know we're on core 22 now (mostly at least?) & that FaH bench 2.3 only covers core21, but I'm surprised not to find a much more recent benchmark thread.
Is that because (as Bruce mentioned on 24/11/19) that F@H uses mixed precision & the benchmark doesn't? So it doesn't reflect actual F@H performance?
But is the mixed precision in F@H varied with each project or is it at a fixed ratio? If it varies, is it by much? If not is it predominantly SP?
If so, then the benchmark still has meaning & would give a near enough comparison to F@H. (Although I understand that a new benchmark is to follow for core22).

[foldy]

Single precision is still the most part of fah work units with a little slower double precision in help where needed.

FahCore_22 has 20% speed acceleration compared to FahCore_21. So results for different GPUs would not be comparable with older FahBench with FahCore_21 anymore.

But for a developer it is easy to build FahBench for FahCore_22 too as it is open source. I can provide a Windows executable if you are interested.

[iceman1992]

FahCore_22 has 20% speed acceleration compared to FahCore_21. So results for different GPUs would not be comparable with older FahBench anymore.

Ah.. and I was using the fahbench results to make a GPU purchasing decision last week
Looks like we can really use a new FahBench version

[_r2w_ben]

FAHBench was mentioned in the Reddit AMA. It's the next priority after the CUDA version of Core 22.

[Assimilator1]

Sounds like foldy has volunteered, so that's good with me, thanks .

[PantherX]

I have finally gotten around to updating the first post with already submitted entries. However, I will no longer be updated since:
1) F@H now uses mixed precision on GPUs but this benchmark doesn't supporting mixed precision
2) FAHBench version 2.3.2 has not been updated to reflect the workings of FahCore_22
3) Over the years, there has has been massive changes in drivers and GPU which means that this list is no longer reflective of real-world. Instead, it is good for historical records.

I look forward to the next version of FAHBench

[Roadpower]

So I'm curious what would long time Folders suggest using while waiting for an updated FAHBench version. There is the perspective of raw performance and the perspective of system stress test. Do you run different packages (more work) to get a more complete idea of what the system is capable of and up to, or do you run one package (less work) and call it a day?

[PantherX]

When it comes to raw performance, shaders and clock speeds are a good indication. However, if it's a new architecture, I am always cautious as historically, a new GPU architecture may take some time to be supported by F@H. Hence I tend to wait for few months to ensure that F@H works, drivers are stable and that there's no hardware issues. Also, changes between architecture might have a positive/negative impact on F@H so that's something I keep an eye out for.

When it comes to stress tests, I have a long list of various benchmarks and stress tests to run in the worst possible ambient temperature. I purposely use a heater to heat up the room where my system is to ensure that I can stress it in the worst possible conditions to figure out the cooling limits. However, with GPU Boost 3.0, I prefer to let the GPU's algorithm do all the clock adjustments for me while I run all the stress tests before hand to ensure it is good enough before leaving it to fold 24/7/365.25 at a normal ambient temperature.

[Roadpower]

Heating up your work space is very dedicated. :p Thanks PantherX