Folding Forum

In the year 2525, if man is still alive ... [/song from my music brain]

I just did a performance comparison using FAHBench with OpenMM 6.2 (Core_21) vs. OpenMM 7.x (future Core_22).

Using Nvidia gtx 1080ti on Windows 7 64bit with FahBench 2.3.1 OpenCL.

Project 13782 Score
OpenMM 6.2: 245
OpenMM 7.2: 1387

Project 9415 Score
OpenMM 6.2: 208
OpenMM 7.2: 1415

So this is a speedup by 6 times.

I did not test AMD GPU yet.

I hope this new core translates into a major increase in the science done, still to be real world tested though.

That's quite impressive ... were you using OpenCL for both OpenMM builds ? Did you try to compile and to compare with a CUDA version of OpenMM 7 ?

Can you share the methodology so that we can try on our cards ?

Hi toTOW, it may be a FAHBench issue were the benchmark simulation runs faster but the real work unit does not.
I would like to test CUDA too but the FAHBench only offers OpenCL or CPU. I guess because the FAhCore does only support OpenCL currently.
So I used only the OpenCL OpenMM for both old and new version and only on Windows, so maybe someone wants to test on Linux too.

Steps to reproduce:
1) Download FahBench http://fahbench.github.io/ and install
2) Create a copy of the FahBench folder and name one FahBench-6.2 and the other FahBench-7.1
3) Download OpenMM 7.1 from SimTK https://simtk.org/frs/?group_id=161 (you need to be registered to download)
4) From the OpenMM zip file extract \lib\OpenMM.dll to FahBench-7.1 folder
5) Delete content of FahBench-7.1\openmm folder and extract \lib\plugins\OpenMMCPU.dll and OpenMMOpenCL.dll to FahBench-7.1\openmm folder

In theory the FahBench-7.1 folder now has a FahCore_21 which runs with the OpenMM-7.1 runtime (aka FahCore_22).

a) Start FahBench-gui.exe from FahBench-6.2 folder and run benchmark on GPU
b) Start FahBench-gui.exe from FahBench-7.1 folder and run benchmark on GPU
[I had problems with the GUI in FAhBench-7.1 so I run both in cmd version.
Get the device and platform id from the GUI first. Then in cmd shell call:
FAHBench-cmd.exe --device-id 0 --platform-id 2 --workunit nav]

You can also copy a real work unit from folding@home to FahBench.
* e.g. C:\ProgramData\FAHClient\work\00\01 copy all files to C:\Program Files\FAHBench\workunits\MyWorkunit
* Copy the wu.json file from another FAHBench\workunits\xxx folder to FAHBench\workunits\MyWorkunit
=> Start FAHBench and now you can select WU MyWorkunit

It was a benchmark issue!

I now could compile the FahBench 2.3.1 from source for OpenMM 7.x

Speed does not improve so much - no magic in new OpenMM 7.x

Sometimes you can replace dlls by a new version and sometimes this doesn't work.

Using Nvidia gtx 1080ti on Windows 7 64bit with FahBench 2.3.1 OpenCL compiled from source for OpenMM 7.x.

Project 9414 Score
OpenMM 6.2: 214
OpenMM 7.1: 258

Project 13782 Score
OpenMM 6.2: 248
OpenMM 7.1: 305

So this is 20% speed improvement in FAHbench OpenCL.

I could run CUDA in FAHbench now too.

Project 9414 Score
OpenMM 7.1 OpenCL: 258
OpenMM 7.2 CUDA: 275

Project 13782 Score
OpenMM 7.1 OpenCL: 305
OpenMM 7.2 CUDA: 324

So CUDA vs OpenCL would be another 6% speedup in FAHbench.

If you are interested I can share my build with you FahBench-2.3.1-OpenMM-7.2.1-with-CUDA-for-Windows.
Just send me a private message.

Still no word on Core 22 or any new core that would take advantage of better hardware?

No news, but the current FahCore 21 works very good on latest nvidia RTX GPUs.

No news doesn't suggest that nobody is work on it. In fact, it's just the opposite. No News is more likely to suggest that more people are working more hours on fixing whatever else needs to be fixed before it's ready for a semi-formal beta test of the new code.

I second foldy's comment. What do you mean "take advantage of better hardware"? What hardware do you have that isn't being taken advantage of when using Core_21?

The main method of taking advantage of new hardware is to update the drivers since that's where hardware features are really supported. FAH talks directly to OpenCL, which either talks to CUDA which talks to the drivers or OpenCL talks directly to the drivers (depending on how your GPU manufacturer installs the drivers).

Looking at the future of computing,
I hope the new core supports and is optimized for more cores.
It seems desktop PCs are getting more power efficient, by running on smaller lithography, allowing higher core count, and higher CPU speeds.
Instruction sets that are ancient are removed and replaced by more updated ones.

GPUs get more cores and slightly higher frequencies.

But PCIE bandwidth remains largely the same.
For gaming there's no need to increase PCIE bandwidth.
So I'm guessing the newer core will find a workaround on the PCIE bandwidth limitation for multiple GPUs?

I'm also interested in the newer qualcomm server chips, their Centriq 2400.
It supposed to have a very high performance per watt, causing many companies to switch from X86 to ARM.

GPUs are halfway ARM. I hope they can port the program to the newest server chips out there.

PCIe 4.0 will double the bandwidth again and be released in 2020 for CPUs and GPUs.

Wasn't Core 22 going to be CUDA? It has been so long ago, I have forgotten.
But if it is still OpenCL, the new AMD Navi chips could be cost-effective.
https://www.digitaltrends.com/computing ... -2070-250/

At least the rumors are impressive.

JimF wrote:Wasn't Core 22 going to be CUDA? It has been so long ago, I have forgotten.

From the small amount of information released a while ago, that is a maybe. Depends on how much extra work would be required to develop and support a second version of the core with CUDA. Also, would depend on whether a CUDA version of the core actually increased throughput and by how much.

Thanks, that refreshes my memory. AMD is showing some signs of life, and whatever works best is OK with me.

Do you think that the future of computing will be graphics cards or CPU?
CPUs are heading towards hexacores and more; it wouldn't surprise me, if a home computer in 10 years, would have between 10 to 50 low-power cores, and inevitably a system that exists just entirely out of 'gpu cores' as CPU cores.
Makes sense from a power perspective.
All the way until quantum computing takes over.