notes on new power mac (was Re: clusters v SMP)
rbw at networkcs.com
Thu Jan 31 06:32:35 PST 2002
Mark Hahn wrote:
> > > yeah, a friend of mine pointed that out. I'm horrified that
> > > Apple's rah-rah-marketing has insideously contaminated my brain!
> > > so basically, if you want near-commodity 64b pointers, you're screwed.
> > > at least if/until AMD does their hammer series...
> > You can get a Sun Blade 100 or Netra X1 for $1000. Does that count?
> well, I suppose it might be possible to want big address space,
> but not give a damn about performance, but that seems like a stretch.
> ie, blade100 delivers a measly 175/182 spec2k int/fp, versus 724/642
> for a cheaper PC clone (AthXP2000).
> it's astonishing to me that Sun even sells these 1998-era machines...
Agreed. Bandwidth to memory is not particularly good either. In all
the design discussions I've seen by Sun engineers on the net (there
aren't many or much detailed info on-line of the USPARC III), most of
the market focus is on low latency to cache which is a figure of merit
of interest to folks with integer codes with dense branching. They
have little to say about floating point and memory bandwidth which
ought to the be focus of Beowulfers. The SpecFP2000 numbers on the
U III 1 GHz look good at first, but are distorted dramatically upward by
one test case.
Sun is much more of a software and marketing company, than a hardware
More information about the Beowulf