What performance does KASY0 really get?
It really gets over 471 GFLOPS on a 32-bit version of HPL. Actually, as of September 16, 2003, we’ve done some additional tuning of the software that allows KASY0 to get 482.6 GFLOPS; our newly tuned SGEMM is being incorporated into Automatically Tuned Linear Algebra Software (ATLAS). Using an “untuned” 64/80-bit version, KASY0 gets a very respectable 187.3 GFLOPS. These aren’t theoretical numbers, they are the real thing. The theoretical we-will-never-see-that numbers are 531 GFLOPS and 1.06 TFLOPS, respectively, for 64/80-bit and 32-bit floating point. Yes, we know HPL is only one application and not a very general one at that. We have other stuff running as well… but most of what we do is computer system design. Thus, our primary applications tend to be things like the Sparse Flat Neighborhood Network design program, which nobody else yet has. The result is that the performance numbers that are most important to us are meaningless to anybody else.
