Optimum processor for MKL / level 3 blass cgemm

[john_shaeffer]

What is the optimum cost effective Intel processor / motherboard to get good MKL performance for level III blas cgemm for matrix sizes in the range of 500 x 500?

Am quite confused as to relative merits of single / dual core, Pentium vs Xenon, 32 vs 64 bit, HT on / off, cache size, etc. for MKL blas 3 performance.

Hardware costs should stay below $4k. The app is LU factoring100,000 unknowns block wise which will also require fast disk IO, thus a MB with SATA II / Raid 0 will be required.

This will be a Fortran 95 (CVF or IVF)code under Win XP.

Thanks.

[TimP]

Dual core clearly improves performance per dollar of parallel applications; that should not be confusing.
Pentium: 1 package per motherboard (2 cores max)
Xeon: > 1 package per motherboard
64-bit mode mainly gives you more addressing and larger arrays; only a small performance improvement where the same application works either way. Were you considering a 32-bit only platform?
HT isn't likely to give a performance improvement in MKL, particularly not in Windows.

[john_shaeffer]

The app will most likely be 32 bit for now, Win XP. That said, is there any benifit to using Linux for MKL orfor disk IO?

How does a single package dual core pentium compare to two Xenon's on a work station board (same clock, cache, etc .) ?

Does cache size make much difference for level 3 cgemm blas?

Thanks.

[TimP]

The answer to the disk I/O question depends on several factors. With Windows supporting RAID disk controllers on fancier Pentium D systems, it should do as well as linux.
In my experience, single package dual core performs as well as dual single core Xeon, when the CPU and bus clock rates are close.
I can't answer your specific question about cache. In general, MKL is well optimized to work with the cache present on any of the targeted Intel CPUs.