rather unfortunate article on Mac
wsb at paralleldata.com
Fri Feb 1 19:18:57 PST 2002
Mark Hahn wrote:
> > did not use that part of the chip. There are prefetch operations that
> > can speed up memory access substantially that go unused. All I've seen
> nah. prefetching will get code up to 80% or so of the peak/marketing
> dram bandwidth. but alas, no Mac has shipped with better than
> 1 GB/s peak dram bandwidth, which is not competitive with the PC world,
> let alone "supercomputer territory".
True, but 800MB/sec is much better than the original posted
results. At 1GB/sec, you're talking about PC133 memory. I'm
surprised the new boxes wouldn't have migrated to DDR PC2100.
> so if your code is "embarassingly cacheable", then rejoice.
> I'd be kind of surprised to find much code that doesn't run
> at least as well on a decent AthXP/DDR box. having GigE builtin
> is sort of amusing, but GigE PCI cards are down to $36 (!).
> the point is that the new Mac is no kind of breakthrough.
Particularly if you need 64bit IEEE...
Although, the specs say:
Powerful floating-point unit supporting
single-cycle, double-precision calculations
Not quite sure what they're up to there. Taking that at face value,
the new chip should be capable of 1Gflop on a 1Ghz processor.
More information about the Beowulf