Re: Consulting about buying new hardware (GPU)

From: michael zhenin (mikelj25_at_gmail.com)
Date: Sun Jul 15 2018 - 06:47:22 CDT

Thanks for your help... If there is no difference between the p100 and the
v100 then there is really no reason to spend more money on the v card.
We tried the TITAN X card and it was slower by ~3 ns/day than the P100
(with Intel Xeon Gold 6130 CPU with 2.10GHz), but because of its price we
wanted to know if the new TITAN V and 1080-TI cards caught up to the P100's
speed.

Thanks again,

Michael.

בתאריך יום ה׳, 12 ביולי 2018, 02:14, מאת Lewis J Martin ‏<
lewis.martin44_at_gmail.com>:

> In my experience GPUs are powerful enough that most workstations will be
> limited by the number of CPUs. If you plan on using NAMD 2.13 when it is
> released then this might change.
>
> Pragmatically it might be better to consider price/performance ratio. P100
> comes with a lot of features that are good for HPC centres, such as ECC,
> but are less useful in a workstation for molecular dynamics. A 1080Ti or
> Titan XP are a fraction of the cost, and you can use the extra money to buy
> more CPUs or GPUs right now or down the track.
>
> From memory I was limited by CPU using a single Titan XP, 20 CPUs (Intel
> Xeon Gold 6138) on a system of ~150,000 atoms with NAMD 2.12.
>
> cheers
>
> lewis
>
> On 11/7/18 4:07 pm, namd-l digest wrote:
>
> Date: Tue, 10 Jul 2018 22:45:27 +0800
> From: JC Gumbart <gumbart_at_physics.gatech.edu> <gumbart_at_physics.gatech.edu>
> Subject: Re: namd-l: Consulting about buying new hardware (GPU)
>
> I just did a test with a 382k-atom system on a p100 and a v100, each with 8 cpus. I get the same performance for NAMD on both (~5.3 ns/day), indicating that itb s CPU-starved. With the p100, the GPU utilization is 33%, while with the v100, its 20%. I doubt that 16 cores would change this comparison enough to differentiate the two.
>
> Best,
> JC
>
>
> On Jul 10, 2018, at 4:44 PM, michael zhenin <mikelj25_at_gmail.com> <mikelj25_at_gmail.com> wrote:
>
> Hi all,
>
> We are interested in buying new GPU card for simulating relatively big systems (more than 300K atoms) using NAMD 2.12 version.
> Currently we have one P100 card (with 16 "real" cpu cores) that is quite fast - around 10 ns/day for that system size.
> We were wondering if there is any recommendation/experience with other cards that will speed up the simulation time. Specifically, V100 PCIE, TITAN-V ,or GTX-1080TI.
>
> Many thanks,
>
> Michael.
>
>

This archive was generated by hypermail 2.1.6 : Fri Dec 06 2019 - 23:19:58 CST