[RC5] re: Teraflops
jpclemen at iwa.dp.intel.com
Fri Mar 17 00:06:09 EST 2000
To respond to some of the things various folks have been saying about
the ASCI Red/Janus/Teraflops machine...
>From Zorba's original post:
> I'm getting curious - how does this compare to dnet?
Oh, probably about 10-12k "bench" (PII 266 MHz) CPU's. The CPU's involved
in the system are the best thing ever put into a socket 8 form factor:
Pentium II Overdrive processors, AKA P6T's. They run at 60/66 external,
300/333 internal, have 512K full speed L2 cache and MMX extensions. I
actually run a few dual processor P6T systems, and they perform a bit
better than two 300 MHz PII's--I never had any 333MHz PII's at my
disposal, but I suspect they're substantially similar at this
problem. FWIW, I believe the ASCI Red is just running them at 300
MHz; power and cooling prevent them from clocking all the way up to 333
MHz, I hear.
Yep, Dnet smashes it; but, as someone else pointed out, it can work on one
problem all by itself. Buying 4,500 dual processor systems would be TONS
cheaper, and, with advances in networking, almost as effective. There are
really few problems that can be broken up among that many processors,
where a distributed computing model wouldn't be a lot more cost
effective. That's what Intel uses for our own internal chip design
purposes: hundreds of networked dual processor systems. Astute
observers will note that Intel has not announced any forthcoming
supercomputer projects. And, yes, as Forbes_Thane at emc.com alluded to,
there are other special, proprietary/one-off things about it that can't be
duplicated by COTS hardware.
>I am kind of disapointed that people in sandia are running pentium
>pros... nothing wrong with that of course, cheap and fast. but I thought
>that they would be using some custom chips or sparcs or mips or something
>in that effect, even crays...
Part of the reason Intel won the bid on that system was our ability to
deliver performance cheaper. The R&D costs on the PPro were spread out
over many, many more units--lots of 1, 2, and 4 processor
systems/servers--so by using "commodity" processors, we were able to
deliver the performance cheaper than the competition. The #2 computer
on the current top500.org list ALSO uses commodity processors.
As far as Pentium Pro chips are concerned, if you check the top500.org
archives from November, 1998, and prior, you see that the "year" was
originally 1997, but in 1999 changed to 1999. Same year, they just
replaced all of the 200 MHz Pentium Pro's with Pentium II Overdrives:
no fundamental architecture changes, just drop-in replacements for over
9,000 processors. Go ahead and try THAT with a non-commodity architecture.
That system has been on the top of the Top500 list for SIX (semi-annual)
iterations; I suspect that its tenure there is nearing its close. Yeah, it
was installed three years ago, when Pentium Pro processors were (well,
close to, at least) state of the art.
Forgive me if I take pride in the accomplishments of my company; I am not
speaking for Intel in any official capacity, and I should have probably
thrown a whole bunch of (TM)'s in various places around processor names.
To unsubscribe, send 'unsubscribe rc5' to majordomo at lists.distributed.net
rc5-digest subscribers replace rc5 with rc5-digest
More information about the rc5