Chris Jones wrote:
> Sorry if this is off-topic, but I was hoping I could draw on the combined
> knowledge of people here.
> 
> Some of the people I'm working for are putting together a distributed
> computing cluster.  As the manager of said cluster, I get to have a say in
> what they buy.  Lots of details are getting sorted out, but one question
> (in various forms) keeps cropping up:  How can we get the most computing
> power per dollar?
> 
> We're primarily interested in floating point, so one suggestion has been a
> bunch of MMX i386's.  But as far as I know, there's no compiler that
> outputs MMX instructions for unix. 

MMX instructions are SIMD (packed) integer instructions, not floating
point.  In fact, FP instructions and MMX instructions cannot be run at the
same time (and switching between the 2 incurs a rather noticeable
penalty).

> I've also looked at the G3 powermacs,
> which are supposed to be fast, but I'm not convinced NetBSD/macppc is
> stable enough yet.

G3 FP performance is not really all that good, perhaps slightly better
than a high-end Pentium II processor.  A fast 604e-based machine has
better FP, I think.

> Other contenders include alpha and Sun boxes, running
> NetBSD or a "native" OS.

I think the SPARC has decent floating point, but the Alphas are certainly
the top as far as things like SPECfp marks go (especially if you get one
of the cryogenically cooled 600+MHz ones)

> Specifics of the machines above, I can get on my own; I'm just wondering
> if there's anything else that I'm missing.  (You know, that 800 GFLOP
> processor that costs 20 bucks.  :)

I don't know of anything that has really good FP performance other than
IBM's POWER3 parts, and I seriously doubt they're all that cheap.

Later.

-- 
Colin Wood                                 cwood@ichips.intel.com
Component Design Engineer - PMD                 Intel Corporation
-----------------------------------------------------------------
I speak only on my own behalf, not for my employer.