Prev: Marketing names (Re: Hinted conditional branch ?)
Next: DoublingStocks - Stock Trading Robot
From: whygee on 18 Apr 2008 21:56
Hi,
thought it'd be interesting to share :

I have been recently invited to a press conference where the french
computer maker BULL announced that they will build
a "300TFLOPS" computer, to be located at the CEA/CCRT
computing center (south of Paris) in 2009.

Here is the (french) PDF of the joint announcement :
ftp://ftp.cea.fr/incoming/y2k01/CCRT/Dossier%20Presse%20CCRT.pdf
See pages 7, 22, 23.

Some arguments hold water but something struck me :
while 103 TFLOPS come from Nehalem XEON processors,
192 TFLOPS come from NVidia GPUs, in SINGLE PRECISION.
I don't know if they count the XEON's TFLOPS in 64 or 32 bits, either.
Counting like that fuzzes the statistics, it's not possible
to compare that to a NEC SX...

Well, the head of CCRT told me that it doesn't matter,
as many applications don't use the full 64-bit precision
of "scientific" numbers. The users compute many different things
and an hybrid machine makes sense for them.
Some of the science applications they run are "explicitely parallel".
Plus, double-precision can be computed with a sequence
of single-precision operations (but you lose the terrific
TFLOPS rating so use the Intel CPUs instead).

Another notable aspect : they say that those 192 "GPU" TFLOPS
represent only 5% of the total acquisition budget.
From what i understand, the rest of the cluster will pump
the rest, with RAM, interconnects, disks... and water cooling !
(all the installed machines today are air-cooled)

Some stats for this machine without a name (yet) :
- 1068 computing nodes with 8 cores each, 16 cores per 1U of rack space
- 48 NVidia computing servers, or 24576 cores (+364 intel cores)
- around 25TB of RAM
- Linux powered \o/

Sure, eveybody is looking at the PetaFLOP milestone in the next few years in Europe.

Comments are welcome to the owner of http://ygdes.com/

yg
From: Chris Thomasson on 19 Apr 2008 01:22
"whygee" <whygee(a)yg.yg> wrote in message
news:48095110$0$21146$7a628cd7(a)news.club-internet.fr...
> Hi,
> thought it'd be interesting to share :
>
> I have been recently invited to a press conference where the french
> computer maker BULL announced that they will build
> a "300TFLOPS" computer, to be located at the CEA/CCRT
> computing center (south of Paris) in 2009.
[...]

Well, GPU can crunch vectors faster than virtually any CPU... IMVHO, it
makes sense to incorporate the two. You can map a vector onto a texture and
tell the GPU to render that sucker. Basic technique of general purpose GPU
programming...

http://www.gpgpu.org

 | 
Pages: 1
Prev: Marketing names (Re: Hinted conditional branch ?)
Next: DoublingStocks - Stock Trading Robot