On 7 Aug 2012 at 21:38, Dave McGuire wrote:
MFLOP/S per dollar? Per watt? Per pound? Per
cubic foot? Per
coolant loop? Vector length? Vector PRESENCE? Parallelism? But
then what is parallelism? (parallel bits, parallel instructions,
parallel processors?)
Allow me to propose a measure--LINPACK. Should run on just about
everything and with excellent pedigree. Further, to make it time-
sensitive, let's define a supercomputer as a machine that has a
LINPACK score at least 75% of that of the most powerful machine
offered by any vendor in the subject machine's year of introduction.
And AFAIK, all supercomputers eat FORTRAN.
The game with supercomputers is to blow away the competition, cost
being very secondary. "Budget" versions with substantially lower
performance don't count. 75% will accommodate near-misses.
So, an ETA-10P doesn't qualify, but an ETA-10 does for 1988. Binary-
compatible all.
A STAR-1B or STAR-65 doesn't qualify, but a STAR-100 does for 1970.
Ditto.
A YMP-EL was not the state of the art for supercomputers in 1992.
133Mflops wasn't state-of-the-art in 1982. Maybe 1972.
Bring it on...
--Chuck