Re: info about gpu hardware

From: Stefano Guglielmo (stefano.guglielmo_at_unito.it)
Date: Sun Mar 10 2019 - 03:38:15 CDT

Thanks for sharing opinion and advice,
Stefano

Il giorno mar 12 feb 2019 alle ore 16:47 Roessler, Bryan C <roessler_at_uab.edu>
ha scritto:

> Multi-GPU simulations are still bottlenecked by the CPU integrator:
> https://youtu.be/TIMHJWmKv4k?t=938
>
>
> I have a hunch that faster DDR4 RAM should give a good price/performance
> return vs additional GPUs (beyond 2) but I have yet to find relevant
> benchmarks.
>
>
> If there is enough interest I may just do the benchmarking myself and
> publish the results online.
> ------------------------------
> *From:* owner-namd-l_at_ks.uiuc.edu <owner-namd-l_at_ks.uiuc.edu> on behalf of
> Vermaas, Joshua <Joshua.Vermaas_at_nrel.gov>
> *Sent:* Thursday, January 31, 2019 1:35:54 PM
> *To:* stefano.guglielmo_at_unito.it; namd-l_at_ks.uiuc.edu
> *Subject:* RE: namd-l: info about gpu hardware
>
> Hi Stefano,
>
> I did a quick test on some nodes I have access to (36 CPUs (2xXeon Gold
> 6154), 2 Tesla V100s), and here are some results I can share.
>
> Small, 36000 atom system:
> 1GPU 115ns/day
> 2GPU 130ns/day
>
> STMV benchmark (1M atoms):
> 1GPU 3.4ns/day
> 2GPU 4.2ns/day
> 1GPU, only half the CPUs: 3.0ns/day
>
> Someone who has done more thorough benchmarking might know more, but there
> are two possibilities for why this might be:
> 1) Contention for PCI-bus resources between the GPUs, since there is only
> so much communication bandwidth available, and architecturally I think
> there is still work that needs to be done each step on the CPU that
> requires the coordinates to be shuttled across the bus.
> 2) Not enough CPUs.
>
> I *think* it is more option 1 than option 2 based on the results if you
> starve STMV for CPUs. Once PCIE 4.0 and its associated parts come out, this
> might get better. If I had to build myself a new machine today, it would
> probably have only one 2070 or 2080, not very much RAM (16GB is more than
> enough for a NAMD compute node that isn't doing anything crazy), and a
> reasonably high clockspeed multicore processor. It kind of depends on your
> budget as to what the exact balance ends up being.
>
> -Josh
>
>
> On 2019-01-31 10:21:05-07:00 owner-namd-l_at_ks.uiuc.edu wrote:
>
> Dear all,
> I am trying to set a new workstation and I would like to know if there is
> a significant improvement in performance with two gpus (gtx 1080 ti or rtx
> 2080) rather than just one, and eventually with which cpu/ram requisite.
> Thanks in advance for any advice and suggestions
> Stefano
>
> --
> Stefano GUGLIELMO PhD
> Assistant Professor of Medicinal Chemistry
> Department of Drug Science and Technology
> Via P. Giuria 9
> 10125 Turin, ITALY
> ph. +39 (0)11 6707178
>
>
> <https://www.avast.com/sig-email?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail> Mail
> priva di virus. www.avast.com
> <https://www.avast.com/sig-email?utm_medium=email&utm_source=link&utm_campaign=sig-email&utm_content=webmail>
>
>

-- 
Stefano GUGLIELMO PhD
Assistant Professor of Medicinal Chemistry
Department of Drug Science and Technology
Via P. Giuria 9
10125 Turin, ITALY
ph. +39 (0)11 6707178

This archive was generated by hypermail 2.1.6 : Sun Sep 22 2019 - 23:20:17 CDT