Greg, It is a bit difficult to use classic apple-to-apples comparison, when there is not much to compare to. vSMPowered systems are the largest x86 systems from both memory and core count, which makes it a bit hard to compare. Some data below:
1. Apples-to-apples: - The fastest SUN 8-proc (x4600m2) AMD machine (2.6): STREAM OMP 9GB - The fastest IBM 8-proc (x3950) Intel machine (3.0): STREAM OMP 4GB - vSMPowered system has linear STREAM bandwidth... so 8-proc STREAM OMP 27GB 2. Apples-to-oranges: - Comparing Itanuim-2 (not 1) and x86 when it get to large memory is the right comparison. After all, finding another x86 system that has >400GB RAM is quite difficult. I don't have SPEC OMP results, but one application I can mention that make use of OMP is Gaussian, which is running by several customers at better performance than other (AMD, Itanuium) platform - after customer's apple-to-apple comparison. I'll be happy to discuss our technology in details if you have the time. Note that we have several contacts in common at PathScale/Qlogic that can share more light on ScaleMP's technology. --Shai -----Original Message----- From: Greg Lindahl [mailto:[EMAIL PROTECTED] Sent: Wednesday, December 19, 2007 13:25 To: Shai Fultheim ([EMAIL PROTECTED]) Cc: [email protected] Subject: Re: [Beowulf] ever heard of ScaleMP? On Sat, Dec 15, 2007 at 10:08:16PM -0500, Shai Fultheim ([EMAIL PROTECTED]) wrote: > - Large memory applications can use all memory - for example > running SANDIA CUBIT with meshes over 400GB in size. 3.08x faster than > the customer large-scale Itanium NUMA system. This is a classic apple-orange comparison. In order to be useful, you'd have to mention what the other system was -- if it's an Itanium-1 at 800 Mhz compared to a 3.0 Ghz modern system, that matters. But, better yet, can you just point us at a set of SPEC OMP benchmark results? I looked at your OEM partner pages and didn't see any benchmark results. -- greg _______________________________________________ Beowulf mailing list, [email protected] To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
