[CP2K-user] [CP2K:14137] Re: Regarding GPU version

[Alfio Lazzaro]

Hi Jörg,
Assuming that you want to run on a single GPU, I think you can fit this 
benchmark:

https://github.com/cp2k/cp2k/blob/master/benchmarks/QS_DM_LS/H2O-dft-ls.NREP2.inp

Note that the problem is not the time, rather the memory available.
Assuming that you can run on a single GPU, then the test should run in a 
reasonable amount of time (minutes), no need to run 1h.
Make sure you use libxsmm for the CPU execution so that you can make a fair 
comparison CPU-GPU executions.

Time ago I gave some explanations on how to understand what the performance 
benefit of the GPU is:

https://github.com/cp2k/cp2k/issues/73

Let me know if you need anything else (or write to me directly, I don't use 
to read the forum so often)...

Alfio





Il giorno giovedì 29 ottobre 2020 alle 17:23:13 UTC+1 sassy ha scritto:

> Hello Alfio,
>
> Many thanks! I am not after the last bit of performance but more to get an 
> idea which of the cards I can use would be better suited for us. 
> If you don't mind: can I run any DFT based job to try that out or are 
> there 
> some which are better suited? I was thinking of one which is running say 
> 1h or 
> so on a CPU. Too short and you might get too much noise, too long and I 
> don't 
> get as much testing done as I would like to. 
>
> I will have a look at the github page. 
>
> All the best from a wet London
>
> Jörg
>
> Am Donnerstag, 29. Oktober 2020, 06:19:21 GMT schrieb Alfio Lazzaro:
> > Hello Jörg,
> > Although CP2K doesn't A100 optimized kernels, you can always use the V100
> > ones. I don't expect a big performance impact.
> > But if you really want to, please check here on how to optimize for
> > A100:
> > 
> https://cp2k.github.io/dbcsr/develop/page/3-developer-guide/3-programming/2
> > -accelerator-backend/2-libsmm_acc/3-tune.html
> > 
> > Best regards,
> > 
> > Alfio
> > 
> > Il giorno mercoledì 28 ottobre 2020 alle 23:26:28 UTC+1 sassy ha scritto:
> > > Hi Fabian, hi all.
> > > 
> > > I am hijacking a bit this thread as I think it is relevant.
> > > I will have access to a GPU test machine which got the new Ampere GPU
> > > cards
> > > installed. If I get the Wiki page correctly, I would need to set the 
> GPU
> > > version to A100, i.e.
> > > --gpu-ver=A100
> > > 
> > > Are the Ampere cards already supported in CP2K? Given that the latest
> > > version
> > > was released *before* the Ampere cards were around, I got the feeling 
> that
> > > might not be the case.
> > > 
> > > Thanks for your help.
> > > 
> > > All the best from a dark and cold London
> > > 
> > > Jörg
> > > 
> > > Am Dienstag, 18. August 2020, 08:41:34 GMT schrieb fa... at gmail.com:
> > > > Hi Bidesh,
> > > > 
> > > > this option will select the which architecture to optimize for. You 
> can
> > > > simply choose the one which is closest to you card. According to
> > > > https://en.wikipedia.org/wiki/CUDA the GTX 1660 Ti has compute
> > > 
> > > capability
> > > 
> > > > 7.5, so you should use --gpu-ver=V100
> > > > 
> > > > Best,
> > > > Fabian
> > > > 
> > > > On Tuesday, 18 August 2020 at 09:23:08 UTC+2 bide... at gmail.com
> > > 
> > > wrote:
> > > > > Hi all,
> > > > > 
> > > > > I don't know if this topic is relevant or not but I am facing a
> > > > > problem
> > > > > while compiling cp2k with cuda.
> > > > > I have a GPU nVIDIA GTX1660ti. I have installed CUDA. Now while
> > > 
> > > compiling
> > > 
> > > > > what should I put in --gpu-ver=. As its showing only K20X, K40, 
> K80,
> > > 
> > > P100,
> > > 
> > > > > V100 are allowed.
> > > > > Thanks in advance.
> > > > > 
> > > > > Bidesh Kirtania
> > > > > Research Scholar
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20201102/698dd989/attachment.htm>