Re: GPU Supercomputer - 300 TFLOPs

On Apr 22, 6:33 pm, whygee <why...@xxxxx> wrote:
AirRaid wrote:
A.) I found a few english articles:
http://www.tgdaily.com/content/view/37070/135/
http://www.tomshardware.com/news/nvidia-graphics-supercomputer,5219.html

thanks !

B.) Nothing wrong with a cross post as long as it's at least somewhat
relevant to the groups posted to. Cross posting can reach a larger
group of people. Though if you want to continue discussion on *only*
comp.arch, that's fine by me ;)

i think it's the best place for this discussion.

C.) I haven't read your post "The next french "supercomputer" but I
want to find it and read it since this is an interesting subject.

sure, it's available on the net on this same NG :-)

D.) The info on this "supercomputer" is interesting because it
apparently lets us know the overall FP performance of Nvidia's new
long-awaited high-end consumer GPU, the GT200. (~1.1 TFLOPs ? )

I have been invited on the site of the future cluster, and it will be
6x more powerful than the existing "platine" supercluster, yet it will be smaller.
(look athttp://www-ccrt.cea.fr/fr/moyen_de_calcul/index.htmfor the current HPC machines).

the performance is announced to be "roughly 300TFLOPS",
consisting of 192TFLOPS in 48 GPGPU servers, or 24K stream processors in 2x 42U bays,
and 103TFLOPS with 1068 Intel Nehalem nodes (not cores or other things...).
Storage and other things (water cooling) will take a lot of room too.
in fact the GPGPU thing is said to cost 5% of the total cost of acquisition
and it takes almost as much space, yet it provides almost 70% of the TFLOPS ...

> E.) What details do you have about the performance of each GPU (GT200) ?
The type of the chips has not been announced.
I'll have to ask Bull's technicians...
The press release is targeted at a wide audience,
like NYT readers... (or the french equivalent).

* GPGPU : 2*12288 stream processors or 192TFLOPS, or 7.8 GFLOPS per processor.
In each bay, some space is used for the Infiniband network.
Then there are 24 GPGPU "servers" coupled to 24 Nehalem servers,
this amounts to 192 Intel cores
BUT the nodes may be dual-chip, so i don't know if the
Nehalem is quad or octo-core.

* the rest of the server uses 19 bays with a density of 16 cores/1U
and it is water cooled.

thanks!

no problem.
i'm working on analysing this announcement
and all constructive comments or informations are welcome.

yg


Thanks for all the details. The main question that comes to mind is
the usability of those GPU-derived FLOPs, for general purpose
programming.
.