[gmx-users] GPU-based workstation

[Szilárd Páll]

Dear all,

As far as I understand, the OP is interested in hardware for *running*
GROMACS 4.6 rather than developing code. or running LINPACK.


To get best performance it is important to use a machine with hardware
balanced for GROMACS' workloads. Too little GPU resources will result
in CPU idling; too much GPU resources will lead to the runs being CPU
or multi-GPU scaling bound and above a certain level GROMACS won't be
able to make use of additional GPUs.

Of course, the balance will depend both on hardware and simulation
settings (mostly the LJ cut-off used).

An additional factor to consider is typical system size. To reach near
peak pair-force throughput on GPUs you typically need >20k-40k
particles/GPU (depends on the architecture) and throughput drops below
these values. Hence, in most cases it is preferred to use fewer and
faster GPUs rather than more.

Without knowing the budgdet and indented use of the machine it is hard
to make suggestions, but I would say for a budget desktop box a
quad-core Intel Ivy Bridge or the top-end AMD Piledriver CPU with a
fast Kepler GTX card (e.g. GTX 680 or GTX 770/780) should work well.
If you're considering dual-socket workstations, I suggest you go with
the higher core-count and higher frequency Intel CPUs (6+ cores >2.2
GHz), otherwise you may not see as much benefit as you would expect
based on the insane price tag (especially if you compare to an i7
3939K or its IVB successor).

Cheers,
--
Szilárd


On Sat, May 25, 2013 at 1:02 PM, lloyd riggs <lloyd.riggs at gmx.ch> wrote:
> More RAM the better, and the best I have seen is 4 GPU work station.  I can
> use/have used 4.  The GPU takes 2 slots though, so a 7-8 PCIe board is
> really 3-4 GPU, except the tyan mentioned (there designed as blades so an 8
> or 10 slot board really holds 8 or 10 GPU's).  There's cooling problems
> though with GPU's, as on a board there packed, so extra cooling things may
> help not blow a GPU, but I would look for good ones (ask around), as its a
> video game market and they go for looks even though its in casing?  The
> external RAM (not onboard GPU RAM) helps if you do a larger sim, but I dont
> know performance wise, the onboard GPU, the more RAM the marrier...so yes,
> normal work stations you can get 4 GPU's for a 300 US$ board, but then the
> price goes way up (3-4000 US$ for an 8-10 gpu board).  RAM ordered abroad is
> also cheep, 8 or 16 MB Vs. Shop...I have used 4 GPU's but only on tests
> software, not Gromacs, so would be nice to see performance...for a small 100
> atom molecule and 500 solvent, using just the CPU I get it to run 5-10
> minutes real  for 1 ns sim, but tried simple large 800 amino, 25,000 solvent
> eq (NVT or NPT) runs and they clock at around 1 hour real for say 50 ps
> eq's....
>
> Stephan
>
> Gesendet: Samstag, 25. Mai 2013 um 07:54 Uhr
> Von: "James Starlight" <jmsstarlight at gmail.com>
> An: "Discussion list for GROMACS users" <gmx-users at gromacs.org>
> Betreff: Re: [gmx-users] GPU-based workstation
> Dear Dr. Watkins!
>
> Thank you for the suggestions!
>
> In the local shops I've found only Core i7 with 6 cores (like Core
> i7-39xx) and 4 cores. Should I obtain much better performance with 6 cores
> than with 4 cores in case of i7 cpu (assuming that I run simulation in
> cpu+gpu mode )?
>
> Also you've mentioned about 4 PCeI MD. Does it means that modern
> work-station could have 4 GPU's in one home-like desktop ? According to my
> current task I suppose that 2 GPU's would be suitable for my simulations
> (assuming that I use typical ASUS MB and 650 Watt power unit). Have
> someone tried to use several GPU's on one workstation ? What attributes of
> MB should be taken into account for best performance on such multi-gpu
> station ?
>
> James
>
> 2013/5/25 lloyd riggs <lloyd.riggs at gmx.ch>
>
>> There's also these, but 1 chip runs 6K US, they can get performance up to
>> 2.3 teraflops per chip though double percission...but have no clue about
>> integration with GPU's...Intell also sells their chips on PCIe cards...but
>> get only about 350 Gflops, and run 1K US$.
>>
>> http://en.wikipedia.org/wiki/Field-programmable_gate_array and vendor
>> http://www.xilinx.com/
>>
>> They can design them though to fit a PCIe slot and run about the same, but
>> still need the board, ram etc...
>>
>> Mostly just to dream about, they say you can order them with radiation
>> shielding as well...so...
>>
>> Stephan Watkins
>>
>> *Gesendet:* Freitag, 24. Mai 2013 um 13:17 Uhr
>> *Von:* "James Starlight" <jmsstarlight at gmail.com>
>> *An:* "Discussion list for GROMACS users" <gmx-users at gromacs.org>
>> *Betreff:* [gmx-users] GPU-based workstation
>> Dear Gromacs Users!
>>
>>
>> I'd like to build new workstation for performing simulation on GPU with
>> Gromacs 4.6 native cuda support.
>> Recently I've used such setup with Core i5 cpu and nvidia 670 GTX video
>> and obtain good performance ( ~ 20 ns\day for typical 60.000 atom system
>> with SD integrator)
>>
>>
>> Now I'd like to build multi-gpu wokstation.
>>
>> My question - How much GPU would give me best performance on the typical
>> home-like workstation. What algorithm of Ncidia GPU integration should I
>> use (e.g SLI etc) ?
>>
>>
>> Thanks for help,
>>
>>
>> James
>> --
>> gmx-users mailing list gmx-users at gromacs.org
>> http://lists.gromacs.org/mailman/listinfo/gmx-users
>> * Please search the archive at
>> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
>> * Please don't post (un)subscribe requests to the list. Use the
>> www interface or send it to gmx-users-request at gromacs.org.
>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>>
>> --
>> gmx-users mailing list gmx-users at gromacs.org
>> http://lists.gromacs.org/mailman/listinfo/gmx-users
>> * Please search the archive at
>> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
>> * Please don't post (un)subscribe requests to the list. Use the
>> www interface or send it to gmx-users-request at gromacs.org.
>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>>
> --
> gmx-users mailing list gmx-users at gromacs.org
> http://lists.gromacs.org/mailman/listinfo/gmx-users
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
> * Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> --
> gmx-users mailing list    gmx-users at gromacs.org
> http://lists.gromacs.org/mailman/listinfo/gmx-users
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
> * Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists