CPU Choices

[Craig Tierney]

On Fri, Sep 21, 2001 at 10:35:25AM -0000, n.gregory at garageflowers.co.uk wrote:
> (sorry if this gets posted twice)
> Hello,
> I am doing some research into the configuration of a 32 1U node Beowulf cluster 
> and have a question regarding CPU configuration.
> 
> The current choice is between Intel P4s as Itaniums seem a little bleeding edge at 
> the moment, or the latest AMD chips.
> 
> AMD seem to be getting a impressive performance for the price, but I?m a little 
> concerned about the lack of mature multiprocessor chipsets and their heat issues.

Have you run your code on an AMD and P4?  Alot of the published
reports show AMD as faster, but mainly for codes that I don't care
about and that havenot been recompiled.

For the codes I am interested in a dual P4 1.7 Ghz is 30% faster
than a dual Athlon 1.2 Ghz when using the Intel compiler for Linux.
SSE support is better in the Intel compiler than the Portland Group 
Compiler.  I didn't seem much improvement using Portland Group and SSE
than non-SSE.  With the Portland Group compiler the two systems perform
similiary (within 5%).

You are going to spend probably 150K (or more) on hardware, but aren't
willing to shell out $1k (or less) for commercial compilers?  For the 
Portland Group compiler, you only need to buy a license for the nodes you
are going to compile on (Front ends).  I do not know about the
Intel compiler, but I hope it is the same.  Anyone know?


Craig

> 
> Intel on the other-hand have the MP chipsets but seem to be falling down with 
> current lack of (non-commercial ) complier that support MMX and SSE, and the 
> whole issue of Rambus Vs DDR memory.
> 
> I would be grateful for any insight into a choice of CPU and its configuration in 
> terms of price/performance/expandability, or any other factor I should be 
> considering.
> 
> Thanks
> Nick
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

-- 
Craig Tierney (ctierney at hpti.com)