Robert Matthew Fesinmeyer writes:
>Has anyone tested a Pentium 4 system using the standard amber benchmarks?
>
>What were your results? (Please include kernel and compiler versions.)
>
>A 1.5GHz system appears to be at ~$2000 now.
I haven't, but I should mention an interesting note. A future version
of gcc, 3.0, which is in development now, will use the SSE2 extensions
for double precision floating point math. It won't necessarily do vector operations,
but using SSE2 for plain old FP is actually much faster than the FP unit on
the P IV. Intel worsened the FP unit on the P IV compared to the PIII, but
they have excellent SPEC benchmarks because they use their own compiler,
which generates vectorized SSE2 code from regular C and fortran.
If you compile your code using gcc 2.95.2, you won't see anywhere near the speeed
you'd expect.
I've been toying with the idea of modifying some of the routines in AMBER to use
hand-built SSE2 code for the expensive operation (1/sqrt(r)) which tends to stall the
instruction pipeline for 100+ clock ticks. This would make igb=3 run much much faster.
The really interesting thing would be to write a nice SSE2-based FFT library, and
have PME use that FFT. I think the speedup would be dramatic. There may already
be FFT implementations using SSE2.
Alternatively, I'd love to hear from a brave soul would uses Intel's reference
fortran compiler to compile AMBER on windows, with an without SSE2 vectorization,
and give us some benchmarks ;-)
Dave
Received on Mon Mar 05 2001 - 11:57:04 PST
