Re: COMP: Physics and computation: two papers

Eugene Leitl (
Thu, 19 Dec 1996 15:46:21 +0100 (MET)

On Wed, 18 Dec 1996 wrote:

> Max More wrote:
> > The completed Intel machine (made entirely from Pentium Pro microprocessors)
> > will be several times faster than the announced result. A 10-teraflops
> > machine is expected by 2001.
> A 5 teraflops machine is expected by 1999, and a 15 teraflops machine
> is expected by 2002. These numbers are the projection of a different
> company (my current Corporate Masters, actually) for a different
> project.

I'd like to see their overall (local/nonlocal) memory bandwidth and their
communication bandwidth. Total TFlops per se are quite pointless, imo.

> Regarding the Intel machines, I'm curious as to the stability and
> robustness of the devices-- how long are they expected to operate
> between scheduled maintenance breaks?

It would be good to see a maspar supercomputer with hot-pluggable
CPU/mem/router modules, and an OS to take advantage of such
paraphernalia. If one snapshot interim simulation stages, one can regress
to the last saved one, so that very little efficiency needs be lost in case
a module dies right in the middle of a step. Using module-RAID redundancy
will be an essential requisite once the number of modules goes much
beyond 1k, or so.

> >The ASCI folks reckon we may ahve a
> > 500-teraflops machine by 2005.
> That strikes me as _slightly_ optimistic, given that I know the 15 TF
> machine by 2002 is pretty much a bleeding edge goal, as of this year.
> But I wouldn't be at all surprised to see (wets thumb, to test the
> wind) 100 TF or so.
> --
> John S. Novak, III
> The Humblest Man on the Net

| |transhumanism >H, cryonics, |
| |nanotechnology, etc. etc. |
| |"deus ex machina, v.0.0.alpha" |
|icbmto:N 48 10'07'' E 011 33'53'' | |