Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Cray

Author: Jeremiah Penery

Date: 19:09:08 07/12/03

Go up one level in this thread


On July 12, 2003 at 17:51:12, Vincent Diepeveen wrote:

>On July 11, 2003 at 12:52:13, Robert Hyatt wrote:
>
>>Different machine.  You won't do this on your Terra:
>>
>>double x [1000][1000][1000];
>>and then do anything useful with it.
>
>that's less than 8GB!
>
>that's indeed not useful for the TERAS.
>
>Too small.

How long will that multiplication take on the TERAS?  Probably a long time.

>It's as i said busy now creating a 8TB database for weather predictions.
>
>To use your way of representing that's a dataset of:
>
>  double x [1024][1024][1024][1024];
>
>Such 8GB arrays you can let 2 oxens handle even :)
>
>350000 cpu hours come down to 350k gflops = 350 tflops = 0.35 petaflops

350 TFLOPs can take about a minute on the Earth Simulator, given only 16%
efficiency.  Weather simulation achieves very good efficiency on the Earth
Simulator (that's its primary purpose), so it's likely to be even faster.

>I guess for matrix multiplications they'll be using for the 8TB database
>something like an approximation library, because a single matrix calculation
>with the weather prediction database is just too slow as you need = 10^40
>calculations for that which even is a bit much for the Earth machine to do :)

I don't know where you get 10^40 from, but you still completely don't understand
vector calculation.

>>That is the power of the Cray, to handle large arrays _and_ stomp through
>>them impossibly quick without regard to NUMA issues.
>
>That's why they're gonna use opterons and hypertransport in the future for
>Crays.

No, they're not.  Cray is building *one* Opteron machine for Sandia Labs.  Go
look up their future plans, involving X1 and X2 machines, which are all *vector*
machines.



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.