Xeon as a worthwhile platform?

Torbjorn Granlund tege@swox.com
11 Apr 2003 11:28:04 +0200


Let me clarify a few things...

Before Opteron gets assembly support in GMP, it isn't going to be
any better than Athlon for running GMP applications.  It might in
fact be slower clock-for-clock.

With well-written 64-bit assembly code for the Opteron,
performance might be as much as 4 times better than in 32-bit
mode.  How does that work?

Consider multiplying two 512-bit numbers.  A 32-bit processor
will need (512/32)^2 = 256 32-bit multiply instructions to form
the product.  A 64-bit processor will need just (512/64)^2 = 64
64-bit multiply instructions.

While we need 1/4 of the number of operations for a bignum
multiply, other factors contribute the the overall GMP speed.
Perhaps a 64-bit multiply has longer latency, making it hard to
keep the loops run at a good speed?  For Opteron, the added
integer registers will allow us to schedule the code for longer
latencies.

I have tried to convince AMD to support a GMP-for-Opteron
optimizatiuon project.  No success thus far.

-- 
Torbjörn

"To accept selfishness as an all-purpose excuse for mistreating others
is to reject the whole idea of right and wrong" - Richard Stallman