mma...@nvidia.com>
From: users On Behalf Of Marzari
Nicola
Sent: Wednesday, July 10, 2019 9:35 AM
To: Quantum ESPRESSO users Forum ; Borelli
Marco
Subject: Re: [QE-users] Bad scaling on GPU version QE
Thanks! Would be great to have inputs/outputs of those benchmarks.
Nicola
Sent from a tiny ke
Dear Michal,
regarding your previous question:
On 7/9/19 9:11 PM, Michal Krompiec wrote:
> Is this implementation restricted to Nvidia GPUs?
Yes, only Nvidia GPUs are currently supported since the accelerated
version is written in CUDAFortran.
Kind regards,
Pietro Bonfa'
--
Pietro Bonfà
9, 2019 9:12 PM
To: Quantum ESPRESSO users Forum
mailto:users@lists.quantum-espresso.org>>
Subject: Re: [QE-users] Bad scaling on GPU version QE
Dear Pietro,
Thank you very much. Is this implementation restricted to Nvidia GPUs?
Best regards,
Michal
On Tue, 9 Jul 2019 at 19:14, Pietro BONFA'
ith only four GPUs enabled).
>
>
>
> Depending on the benchmark I normally see 6x to 11x speedup.
>
>
>
> Best regards,
>
> Miloš
>
>
>
> *From:* users *On Behalf Of *Michal
> Krompiec
> *Sent:* Tuesday, July 9, 2019 9:12 PM
> *To:* Quantum ESPRESSO u
, July 9, 2019 9:12 PM
To: Quantum ESPRESSO users Forum
Subject: Re: [QE-users] Bad scaling on GPU version QE
Dear Pietro,
Thank you very much. Is this implementation restricted to Nvidia GPUs?
Best regards,
Michal
On Tue, 9 Jul 2019 at 19:14, Pietro BONFA'
mailto:pietro.bo...@unipr.it>> wrote
Dear Pietro,
Thank you very much. Is this implementation restricted to Nvidia GPUs?
Best regards,
Michal
On Tue, 9 Jul 2019 at 19:14, Pietro BONFA' wrote:
> Dear Michal,
>
> I will hopefully be able to share some new benchmarking results in the
> wiki soon, but for the time being the most
Dear Michal,
I will hopefully be able to share some new benchmarking results in the
wiki soon, but for the time being the most accurate estimate is in this
presentation:
https://on-demand.gputechconf.com/gtc-eu/2018/pdf/e8340-porting-quantum-espresso-to-gpu-accelerated-systems.pdf
.
In the
Dear Pietro,
What is the typical speed up vs a cpu-only system? Is this
https://www.dcs.warwick.ac.uk/pmbs/pmbs17/PMBS17/pres/paper3.pdf still
valid? Can you share any benchmarks on V100?
Best,
Michal Krompiec
Merck KGaA, Darmstadt, Germany
On Mon, 8 Jul 2019 at 20:29, Pietro BONFA' wrote:
>
Dear Jing,
good point, I'll add the list of accelerated features in the wiki. In
the mean time you can check page 21 of the first presentation that you
can find here: https://gitlab.com/QEF/q-e-gpu/wikis/home#performance
Exact exchange will come with the next release but you can already try
it
I am working on the latest version 6.4-a1. I have followed the instructions
on the link you provided during the installation. I would like to know
which part of the calculation in pw being accelerated by GPU. For example,
is it boosting the matrix diagonalization during scf cycles, or boosting
the
Dear Jing,
which GPU-enabled version are you using?
GPU support starting from v6.1 is only available when using PGI
compilers (well, in principle, any compiler implementing CUDA Fortran).
You can find additional information here:
https://gitlab.com/QEF/q-e-gpu/wikis/home
Kinds regards,
Pietro
Hi,
I am trying to test the scaling of GPU version of QE recently. I found
out the computational time scaling of the GPU version is very bad. Is there
any unique flags or keywords I don't know about the GPU version? I am using
gcc-6.2.0, openmpi-1/8/4, craype-accel-nvidia35. I know it is not
12 matches
Mail list logo
