When Intel went to the P4 they tossed the dedicated math core, so they
could devote more chip area to the pablum that the general public
wants.
The whole AMD vs. Intel argument isn't really 'interesting' to me,
but one thing about the IntelP4 -- it does tend to underperform
when using the old stack-style instructions for FP math. When using
the SSE/SSE2 instruction set, it performs much better, but with a slight
hit WRT math accuracy. My application includes FIR filtering, some
transforms and some matrix operations.
I have been writing significant amounts of FP code with the SSE
instruction set, and the P4 is very very fast. It is important
to take advantage of the SIMD capabilities (which isn't always possible)
instead of limping along with the 'stack' instructions. (Of course, the
stack instructions can give higher precision or give access to the
microcoded transcendental math at high accuracy.)
When doing DP math when using the stack instructions (which is what
most compilers tend to use) and comparing the SSE2 SIMD instructions,
based upon real code, the P4 can blow itself away (speed wise) in SSE2 mode.
2X performance improvement for vector (or matrix) operations for SSE2
operations isn't unrealistic (but that much isn't guaranteed.)
The SSE2 instructions do their best for multiple (SIMD) operations -- things
like array operations, which are great for DSP, vector and other
operations.
I'd still like to have an SMP FX53 Athlon64, mostly for the potential
for improved performance for sundry things (in 64 bit mode, some operations
appear to be very quick.) Normally,
I prefer dual processor machines, and am disappointed that upgrading
(or side-grading) to AMD64 from Intel would cause me to loose the ability
to simulate SMP multithreading when choosing the highest performing chips.
(My purposes for using SMP isn't always for max performance -- even though
the extra performance can be nice, but it is the ability to check out
code for real SMP machines... So, HTT helps to give me some of what I
want for SMP, even though it doesn't give as much extra performance as
SMP.) I'd happily 'upgrade' from IntelHTT to Athlon64 SMP, but would
loose the choice of the fastest machines for SMP (or SMP simulation.)
John
