FAHBench not working

FAHBench not working

This is a Folding @Home work in progress Benchmark. Not much to say about this. Tried different drivers but no luck. 

http://fahbench.com/

27 posts / 0 nouveau(x)
Dernière contribution
Reportez-vous à notre Notice d'optimisation pour plus d'informations sur les choix et l'optimisation des performances dans les produits logiciels Intel.

We are aware of and working on fixing some OpenMM issues. Can you provide more details on what exactly is not working?

Thanks,

Raghu

There are different behaviors with different drivers stacks. For 15.31.64.2885 I will see what error I get.

FAHBench.exe -deviceId 0 -platform OpenCL -precision single

After that command program says: checking for accuracy...done

And then it immediately crashes (appcrash FAHBench.exe).

15.26.13.64.2778:

Checking for accuracy...Exception: <Force Group -1 >. Potential energy error of 54295.9, treshold of 33.411

Reference Energy: -334110 Platform Energy: -279814

15.28.13.64.2963:

clSetKernelArg

Sorry for not getting back to you sooner. I tried an internal driver build and I get

Explicit:
Checking for accuracy...done
3.27851 ns/day
Implicit:
Checking for accuracy...done
18.6941 ns/day

 

Not sure what the expected output is but I cannot reproduce what you are seeing. I'll install the versions you mentioned and try again and get back to you.

Thanks,

Raghu

That looks nice and performance pretty good, comparable to A10-5800K or GT430. It is fixed already then it seems. When does Intel release the driver?

We dont have a date yet, but it'll be out shortly. I'll let you know as soon as the driver is out.

Thanks,

Raghu

I was able to get it to run on our IVB CPU and GPU with the same driver that Raghu is using. Here are the results:

CPU => Explicit: 3.29623 ns/day, Implicit: 18.7814 ns/day

GPU => Explicit: 3.10708 ns/day, Implicit: 3.13564 ns/day

Do those numbers look good? Also we noted some strange behavior with the benchmark. We found that you could put any number in for the deviceId and it would still run. Is this the expected behavior?

I'm not the developer, can't answer that question. There is a dev mail address so you might write them a mail and ask. As for the numbers, yes they are pretty good for HD4000. Your 18.7814 ns/day result must be from GPU not CPU. A 4 core CPU can't be that fast.

Here you have some for comparison: http://foldingforum.org/viewtopic.php?f=38&t=23440

I noticed there is a HD4000 result on the last page. Comparable to A10-5800k (7660D) and GT430, usually we would expect 7660D and GT430 to be much faster than HD4000 but not the case here.

I could check again, but I used task manager and GPA to ensure my tests were being ran on the appropriate devices. I don't have access to your sources and there was the issue with the binaries accepting various deviceIds. There may be some other issues as to why the numbers came out the way thay did. I ran the tests several times and the results were fairly consistent. I could try on a different machine though? 

You might check GPU load with GPUz while running fahbench. If it runs on GPU then GPU load should be something around 100% I guess. Did you run the benchmark with "FAHBench.exe -deviceId 0 -platform OpenCL -precision single"?

I could try it myself finally.

CPU 3570k

Explicit:
Checking for accuracy...done
2.60685 ns/day
Implicit:
Checking for accuracy...done
2.69118 ns/day

GPU HD4000 (3570k)

Explicit:
Checking for accuracy...done
3.34347 ns/day
Implicit:
Checking for accuracy...done
18.6256 ns/day

I am glad it worked for you.

Thanks,

Raghu

Hi Michael,

I wrote FAHBench (I also work on OpenMM and Folding@Home). Thanks a lot for testing this out, I'm quite happy to see Intel working on their OpenCL implementation. 

As for Chuck De Sylva (Intel)'s deviceId question, FAHBench is designed to allow it to work on multi-GPU/CPUs as well. Version 0.5 didn't do a check for validity of indices, the next version will have better checking. Valid deviceId should be reported by the --display-devices. A lot of us run on systems that have ridiculously combinations of devices that support OpenCL. 

PS - we'd love to hear more about the Xeon Phi's

http://semiaccurate.com/2013/04/29/a-look-at-intels-opencl-performance/

Performance wise, Intel appears to have a rather mature OpenCL implementation which is surprising considering that Intel has just added OpenCL 1.2 support whereas AMD has been offering that feature for quite some time.

But their Explicit score for the i7-3770k is too low. 3.4 is the correct score.

FAHBench 1.2.0 doesn't work anymore with HD4000. Not good.

Here with my code (completely unrelated to FAHBench) I have encountered a bug in the latest driver and the one before that.

A fix (been told that they had it fixed on a branch) might make it to the next driver but until then I am stuck with 9.17.10.2932.

FAHBench 1.2.0 might be a similar issue.

Michael,

Can you provide some details and steps to reproduce the issue you are seeing. I will take a look and see if its the same issue that Laurent was seeing with his application.

Thanks,
Raghu

Start the benchmark on HD4000, programm says benchmarking and nothing happens. The system freezes and I have to reset my PC.

Ok that doesn't sound like my issue at all :) At least I was able to run my program fine, the result wasn't correct though.

What version of the driver were you using when your system froze?

 

I tested 9.18.10.3071 and 9.18.10.3131. I heard that running on CPU works fine (haven't tried myself). Only the GPU doesn't work.

Matches what I saw with my code too. I had to revert to the driver before these two.

On the compute side, I couldn't get our Folding@Home benchmark to work

http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-c...

I'm not the only one.

Michael,

Apparently FAHBench is not supposed to be run on Intel Graphics (coming from OpenMM and FAHBench developers). This is a known issue and the default behavior of the benchmark is to run on CPU when running on Intel Platforms with processor graphics. Did you try forcing the benchmark to run on the Intel GPU device?

Raghu

I didn't try CPU. They told me this is a driver issue from Intel and I reported back here.

Laisser un commentaire

Veuillez ouvrir une session pour ajouter un commentaire. Pas encore membre ? Rejoignez-nous dès aujourd’hui