I've been updating the information on tuned BLAS for R-admin in R-patched
and R-devel. We have
ATLAS (widely available, including for Windows)
MKL(licensed on ix86 and x86_64 Linux and Windows)
ACML(by AMD, but for all ix86 and x86_64 chips, Linux and Windows.
Now available for gfortran.)
Goto(academic use only, only some chips, only Linux)
MKL and ACML provide full LAPACK, the other two some optimized LAPACK
routines. (We have an MKL licence with our icc/ifort licences but it has
not been delivered yet so I used a non-commercial Linux-only download.
Hence I have not tried Windows.)
32-bit Linux I used my dual Athlon 2600 MP desktop (about to be
replaced). Goto no longer supports that chip, and ACML is not threaded
(for gcc). ACML was a little faster than ATLAS, which was faster than
MKL. However, MKL exploited the two processors to halve the elapsed time.
MKL on that chip is poor on complex linear algehra.
64-bit Linux I used a dual Athlon 248. Here the Goto BLAS was the
fastest, but only just faster than ACML when using one CPU. ATLAS was
slightly slower, and MKL perhaps 20% slower but good at exploiting 2
CPUs. This time it was not relatively slower at complex algebra.
Windows ACML is effective. I tested my laptop, a 2GHz Pentium M
(such chips are far faster than their GHz would suggest).
ACML outperformed ATLAS by 10-25%.
These comparisons are biased as I have not compared MKL on Intel
processors. That's lack of interest as all our current compute servers
are AMD.
The revelation was ACML: fast, easy to use even on Windows and completely
gcc-compatible.