There is a debate going on in my company. FEM people say that PENTIUM III is much faster and best suited for scientific application whereas CFD people say that SPARC is much stable and preferable! Who is right and why? Is there any web site where bechmarks and/or detailed informations on which is better are available? Thak you, lawrence

The Pentium architecture is badly suited to numerical work being crippled by waiting for data to come from memory. This was always the major difference between proper computers and microprocessor based computers in the days when there were proper computers. I would estimate that for a large numerical problem a typical PC is spending less than 5% of the time doing useful computatation and 95% of the time heating the air.

Here is a fairly recent thread broadly on this subject showing a SPARC to be significantly more efficient:

http://www.cfd-online.com/Forum/main...cgi?read=16606

it didn't really get anywhere in terms of solid facts (and those presented need careful interpretation!).

However, relatively few large clusters are built using SPARC processors. A quick look at the prices Sun advertise for the servers may well reveal why! In terms of number crunching ability the Pentium-type processor appear to be more cost effective even though you need more of them.

I'll add that the intel memory subsystem and cache access is very very inefficient when compared to unix based workstations with similar SPECmarks (this is not limited to Sun's stuff. The same is true of other manufacturers of UNIX workstations).

The situation has gotten somewhat better lately with the cache prefetch instructions available on both Athlon's and Pentiums. Compilers have only just become available in the last 6 months which can take advantage of this feature.

Dan.