[Beowulf] CCL:Question regarding Mac G5 performance (fwd from mmccallum at pacific.edu)
hahn at physics.mcmaster.ca
Fri May 21 06:14:09 PDT 2004
> The only benchmarks for rx1600 are at 1 Ghz, and SPECfp peak is only
> 1382. If the $3,000 system is only 1 Ghz, that's not so great price
> performance, compared to say a 2.2 Ghz Opteron with a SPECfp peak of
right. I did a little meta-analysis of specFP recently, trying to get at
the fact that some machines owe much of their score to a small number of
spec components. I simply sorted the individual spec scores for a machine,
and recomputed specFP omitting the top 1, top 2, etc.
as a user, what you'd like to see is a long, flat section in the middle.
I thought a couple things were interesting about this. first, for It2,
the high scores correspond with components that have tiny working sets,
and fit in cache. you can also see this reflected in how the two it2/1300
the other interesting thing is that the opteron (148, I think) is actually
faster than the it2/1300 for a significant subset of specFP.
that said, it seems pretty clear that:
- ppc970's are great if your code is cache-friendly and can launch
two vector mul-add's per cycle.
- it2 is great if you can afford them and your code is cache-friendly
and your code is amenable to the kind of pipelining required by EPIC.
- opterons are great if your working set makes caches less effective.
of course, most installations will also need to consider power - though no
one seems to know how much the 90 nm ppc970 dissipates...
More information about the Beowulf