> >> Results of the tests revealed that the SGI Altix 330 using shared > >> memory and unmatched I/O throughput achieved superior > >> performance and ran faster than twice the number of cores in > >> comparable AMD dual core Opteron-based systems. well, it's a marketing paper: it's not in SGI's interest to explore the whole vast, many-dimensional configuration space to find the absolute best way to configure Opteron systems. there's no doubt It2's run some code very well, and that SGI's NUMAflex is a very nice interconnect. but any honest reader must wonder: - what would the results be for 8 processors? (that is, higher-clocked Opterons, and not dual-core to avoid memory contention.) - what are the relative prices of the systems tested? - was the Opteron configured with 8 banks of PC3200? - why was sles9 used rather than sles10? - why was gcc 3.3.3 used rather than 4.x or even 3.4.5? - is mpich-1.2.7-ch_shmem really an optimal choice? - which version of PGI was used on opterons? switches? - for the Opteron benchmarks compiled with Intel compilers, was it running actual SSE3 codepaths, or defaulting to PIII? - is Intel LAPACK/MKL better or worse on Opterons than ACML/etc? regards, mark hahn.