AMBER is designed to run the entire simulation in GPU-space. The CPU acts
as a glorified foreman shouting out directions to the GPU as to which tasks
it should run. If AMBER were rewritten from the ground up, that role could
be removed entirely. The upshot is that for single-GPU runs, you could
probably stuff 8 GPUs into a single box and barely notice the difference.
And you could also use a cheap consumer CPU as well since it's mostly
sitting on its butt waiting on the GPU(s).
The difference between using a 4 year-old $20 CPU and a current $1000+ CPU
in this situation is about 2-3%.
2011/9/13 Marek Maly <marek.maly.ujep.cz>
> Hi Ross,
>
> first of all thanks a lot for your complex answer !
> In fact I assume mainly independent single GPU jobs. So if I understood
> well, in such case there should not be problem considering below mentioned
> motherboard/CPU/4xGPU. Am I right ?
>
> Best wishes,
>
> Marek
>
>
>
>
> Dne Tue, 13 Sep 2011 18:53:34 +0200 Ross Walker <ross.rosswalker.co.uk>
> napsal/-a:
>
> > Hi Marek,
> >
> > I would say that the answer is, with caveats, not really. The reason for
> > this is not a core to GPU ratio argument per se but actually one of
> > memory
> > bandwidth. For a single socket motherboard there is no way that it can
> > drive
> > 4 PCI-E sockets flat out at x16 speed. Hence you are always going to get
> > contention. Typically I would say never go beyond 2 GPUs per node for a
> > single socket system. 4 GPUs per node is pushing it for a dual socket
> > system
> > even with the highest CPU FSB speed you can get so on a single socket
> > system
> > with a low end CPU it is definitely not going to cut it.
> >
> > Now if you only plan to run single GPU jobs. I.e. 4 independent jobs at
> > once
> > with reasonably large values for ntpr, ntwx etc then it is probably not
> > an
> > issue since the memory bandwidth only really comes into play at each
> > output
> > file or trajectory write for most types of calculations. However, trying
> > to
> > run in parallel where a single job spans GPUs it really all comes down
> > to a
> > combination of the aggregate memory bandwidth of the main CPU memory per
> > GPU
> > and the PCI-E bandwidth to each GPU when all are communicating flat out
> > at
> > the same time. You might be able to get speedup across 2 of the GPUs if
> > you
> > leave the other two idle. However, running two sets of dual GPU runs or
> > running a single 4 GPU run is likely to cause so much contention that see
> > little or no speedup.
> >
> > So it really comes down to a case of the type of jobs you want to run.
> >
> > I hope that helps.
> >
> > All the best
> > Ross
> >
> >> -----Original Message-----
> >> From: Marek Maly [mailto:marek.maly.ujep.cz]
> >> Sent: Tuesday, September 13, 2011 7:19 AM
> >> To: amber.ambermd.org
> >> Subject: [AMBER] Sufficient CPU cores/GPU ratio ?
> >>
> >> Hi all,
> >>
> >> just a very quick technical question. I would like to know if 4 core
> >> CPU
> >> "Intel Core i7-960" might be sufficient for
> >> 4 GPU machine (motherboard: Asus P6T7 WS SuperComputer - Intel X58 ).
> >> I
> >> mean if ratio 1 CPU core/ 1 GPU might
> >> be here sufficient also considering some requirements (on CPU) for
> >> managing operating system etc.
> >>
> >> Thanks in advance for any relevant comment !
> >>
> >> Best wishes,
> >>
> >> Marek
> >>
> >> _______________________________________________
> >> AMBER mailing list
> >> AMBER.ambermd.org
> >> http://lists.ambermd.org/mailman/listinfo/amber
> >
> >
> > _______________________________________________
> > AMBER mailing list
> > AMBER.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber
> >
> > __________ Informace od ESET NOD32 Antivirus, verze databaze 6459
> > (20110913) __________
> >
> > Tuto zpravu proveril ESET NOD32 Antivirus.
> >
> > http://www.eset.cz
> >
> >
> >
>
>
> --
> Tato zpráva byla vytvořena převratným poštovním klientem Opery:
> http://www.opera.com/mail/
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Sep 13 2011 - 11:30:06 PDT
