[Tfug] AMD vs. Intel floating point

[John Gruenenfelder]

On Sun, Mar 30, 2003 at 03:23:56PM -0700, Anthony Steckman wrote:
>
>Many of us at TCS run the Folding@home distributed computing software 
>and it has been shown fairly consistently that an AMD will outperform an 
>Intel of equal speed in terms of work unit output. People who seem to 
>know about these things always say it's because of the way the AMD's 
>handle floating point calculations.
>
>I wouldn't stand by the assertion -- I hardly know what I'm asserting. I 
>can say it's at least a very well established belief in the Folding@home 
>community.

Interesting.  However, you say that is of chips at equal speeds?  In this case
my choices are an Athlon MP 2700 and a P4 3GHz.  As far as clock speed, the
AMD chips is 2/3 that of the Intel chip.  Comparing "speed numbers", the AMD
chip is only 400 lesss than the Intel.

Another thing I should have mentioned is that, unlike many of the distributed
computing jobs, our jobs are very much memory bound.  That is, they are very
big.  On our dual P3-800, we have seen that starting two jobs will decrease
the speed of each by 30% comparred to the speed they would run at if run
alone.  This is still faster than running them serially, but it shows much
memory contention.  I have read in a number of places that the AMD set up has
much better performance over the Intel MPS  (muilti-processor specification).

Given that, perhaps even if the 2700 were a little slower than a P4-3GHz, the
increase in memory speed (less contention) would make up for it?


-- 
--John Gruenenfelder    Research Assistant, Steward Observatory, U of Arizona
                        johng@as.arizona.edu
Try Weasel Reader for PalmOS  --  http://gutenpalm.sf.net
"This is the most fun I've had without being drenched in the blood
of my enemies!"
        --Sam of Sam & Max