[Beowulf] HPC with CUDA

[rf at q-leap.de]

>>>>> "Prentice" == Prentice Bisbal <prentice.bisbal at rutgers.edu> writes:

Hi Prentice, Raphael,

    Prentice> Raphael, Not many vendors make servers that support that
    Prentice> many GPUs is a single chassis. I know for a fact Dell
    Prentice> doesn't. I think the HP Proliant servers are your best
    Prentice> option.

definitely check out Supermicros offerings as well
(http://www.supermicro.nl/products/nfo/gpu.cfm). They make really great HPC
hardware and have tons of options that should suit your
requirements. It's our preferred hardware for Qlustar clusters. You can
buy them from many resellers.

Disclaimer: We have no affiliation with Supermicro whatsoever. Just
happy users ... 

-- 
Roland

-------
http://www.q-leap.com / http://qlustar.com

    Prentice> Processor shouldn't be to important in terms of
    Prentice> compatibility with the GPU. If all the work is going to be
    Prentice> done on the GPUs, you could probably save some money
    Prentice> here. If there's going to be a mix of GPU/CPU work, I'd
    Prentice> shoot for whatever processor is currently at the pricing
    Prentice> 'sweet spot'.  Talk to your vendor to find out which
    Prentice> processor(s) are currently in the sweet spot for pricing.

    Prentice> For RAM, I recommend you have at a minimum least as much
    Prentice> main RAM as you have RAM in all of your GPUs
    Prentice> combined. Since the GTX 780 has 3 GB of RAM, you would
    Prentice> need a minimum of 15 GB of RAM on the server that will
    Prentice> have the 5 GTX 780s GPUs. That's not really a lot of RAM
    Prentice> these days. I think it would be better to have at least
    Prentice> double the amount of GPU RAM for main ram, so I'd
    Prentice> recommend at least 32 GB of RAM in the server with 5 GPUs,
    Prentice> which is still not a lot of RAM by today's standards.
    Prentice> With today's multicore processors and 2-4 GB RAM per core
    Prentice> as a rule of thumb, you are probably already exceeding
    Prentice> those RAM recommendations.

    Prentice> An important thing to look into is how many PCI buses are
    Prentice> in your server. For the server with 5 GPUs, you'll want
    Prentice> more than one and you'll want to distribute your GPUs as
    Prentice> evenly as you can across the busses to reduce contention
    Prentice> on the PCI bus.

    Prentice> The most important issue is to make sure the server has
    Prentice> adequate power and cooling capacity. GPUs use a lot of
    Prentice> power and give off a lot of heat. Without adequate power
    Prentice> and cooling within the server chassis, you might see some
    Prentice> flakey behavior as voltage dips or temperature increases
    Prentice> too much.

    Prentice> Prentice


    Prentice> On 06/03/2014 09:51 PM, Raphael Verdugo P. wrote:
    >> Hi,
    >>
    >> I need install 5 GPUs(Geforce GTX 780s) in a server and 1 Tesla
    >> Keppler K40 in other.
    >>
    >> ¿ Do they have any recommendation for server HP or Dell?
    >> processor? , RAM?
    >>
    >> I review a HP ProLiant DL380p Gen8 Server (1) . Do they any
    >> observation?
    >>
    >> We will use it with Nvidia CUDA for HPC...
    >>
    >> We have the GPUs, I need a server for install it.
    >>