speed?

dear R wizards:

while extolling the virtues of R, one of my young econometrics
colleagues told me that he still wants to run ox because [a] his code
is written in it (good reason); [b] because ox seems to be faster than
R in most benchmarks (huh?).

this got me to wonder.  language speed can't matter much, so it must
be mostly the underlying matrix algebra by now.  I presume that
nowadays most of the plain matrix operation speed depends primarily on
which hardware features the library accesses.  (The basic algorithms
aren't so different, so even though the algorithm may have mattered a
long time ago, they are probably pretty similar nowadays. hmmm...maybe
matrix inversion still is different, but multiplication and adding
should not be.)

On x86 architecture, I believe there is a hierarchy in terms of raw
processing power:  FPU < SSE* < GPU.

is it even possible to use the GPU now for math processing (linux or
windows), and specifically in R?

assuming I compile everything with the proper SSE flags and atlas, is
SSE* fully taken advantage of?

regards,

/ivo