On Thu, Jan 22, 2004 at 10:21:24AM -0500, Peter Shenkin wrote:
> Would you ever expect something like a 30% improvement for
> Opteron, based on special compilation?
No. For this example, vectorized floating point instructions provide
little benefit on the Opteron, because the scalar ones are so fast.
One of our guys wasted a couple of weeks doing FP SIMD before we
learned this one. SIMD is a benefit for vectorizable integer
operations.
I'd recommend approaching your question from a different angle: Look
at SPEC results for Opteron and Pentium, and compare peak performance
with base performance compiled with the flags from the other cpu. That
is, peak Opteron performance as published against Opteron performance
using the base Pentium flags. You'll have to run the second number
yourself.
-- greg
|