> the largest Origin system that SGI has announced is only 128 processors
You are correct, but at that size, the only 6 computers sold in the world today are faster as supercomputers. (based on an approximation based an anticipated LINPACK results, once it uses all 128 processors) (also please note that LINPACK performance is not a good measure of integer or bandwidth performance. It's a big matrix inversion, a task that splits up onto many processors well. So just because you have a great linpack number doesn't mean you've got a great server, or I/O or anything else. Linpack is however a **useful** HPC figure as **many** supercomputing tasks involve at some stage a big matrix inversion or a big "time step iteration over a large number of cell" that have similar performance characteristics as a large matrix inversion)
Intel ASCI Red
a one off for the gov't experimental computer
SGI/Cray T3E...
various models, its the fastest commerically available.
Hitachi/Tsukuba CP-PACS/2048
apparent a one off at a Japanese supercomputing lab
Hitachi SR2201/1024
commericially available 1 at this size exists, 10 total
of this type are in the top 500 list. Only one outside Japan
Fujitsu Numerical Wind Tunnel
another one off for a very specific purpose
Fujitsu vpp (several models
a fast vector machine not well suit for the server role.
IBM SP (several models, some faster, some slower)
a nice fast massively parallel machine which is apparently a
nightmare to program for optimal performance
Intel XP/
a discontinued Intel supercomputer
NEC SX (several models, some faster, some slower vs. 128proc 02K)
don't know much about this one. Installed base -- 29 of the
top 500 -- 4 outside Japan.
SGI/Cray T3D
predecessor to the T3E
Hewlett-Packard Exemplar X-Class (several models, some faster, some slower vs. 128proc 02K)
26 of the top 500 sites world wide. Pain in the neck
mulitprocess programming model. (from what several PhD types
have told me)
Thats 6 commercially available, only two of which could be considered "server" competition (the HP and the NEC SX). The TOTAL number of reported systems in the world faster than an Origin 2000 128proc (assuming as yet unannounce performance number -- okay so it's a WAG)
about 90. Again that's 90 total installed computers in the entire world faster than a Origin 200 128proc.
What's really cool, and really mind-blowing, is that SGI set one up (in apparently less than two days) on the trade show floor of "SuperComputing '97" this week. The words "portable" and "supercomputer" don't usually go together -- but the ability to put together a top 100 (in the world) computing site in a generic space, in a short period of time is astounding.
Assumptions and references...
for actual Top 500 results:
netlib.org
The approximate estimate of 128 proc Origin 2000 results is based on an assumption of 80% efficient scaling -- i.e. 1.8 * 64 processor results (for an approximate RMax of 34,200 -- which would rank it at 87th) , which is consistant with the 16-32, an 32-64 ratio and the parallizability of the problem. Actual results (when they are released) could be better or worse, YMMV, (insert disclaimer here).
Very unofficially,
john |
