This claim deals with the maximum megaflops (MFlops) rate achieved by any computer on the n=1000 LINPACK benchmark, as published in any issue of "Performance of Various Computers Using Standard Linear Equations Software" by Jack J. Dongarra with a publication date on or before 31 December, 1999. The same second paragraph as LA25 and LA30 applies here. YES coupons will be worth $1.00 if the maximum published rate is at or above 600,000 MFlops. YES coupons will be worth $0.00 if the maximum published rate is below 100,000 MFlops. If the maximum published rate is between those values, the value of a YES coupon will be (MFlops - 100,000) / 5,000. The current peak rate is approximately 20,800 MFlops as of January 1995. 600,000 MFlops would require a sustained 2x per yer increase. 100,000 MFlops would require a sustained 1.3x per year increase. An ongoing U.S. effort called the "Grand Challenge" has the goal of producing a machine capable of a teraflop (approx. 1,000,000 MFlops, depending on how exactly you define it.)
I will judge based on the wording of the claim unless it is found to be ambiguous. Such ambiguities will be resolved based on my perception of the author's intent.
The "second paragraph" referred to above reads "If the use of the n=1000 LINPACK benchmark is discontinued, published performance on any essentially identical benchmark will be acceptable. If the series of papers of the above name is discontinued, any paper authoritatively detailing performance on an acceptable benchmark will be acceptable."
The newest version of Dongarra's paper is available online in postscript form.
Just to be perfectly clear, the current version of that paper doesn't have a "LINPACK n=1000" benchmark. The closest benchmark is the "TPP Best Effort n=1000" benchmarks, which is based on LINPACK but allows the manufacturer to substitute a hand-optimized implementation. I suspect that benchmark is the one the author intended to cite. I will base the judgement on that benchmark if it is still in use when the claim is to be judged. The best results reported in the current (May 7 1997) version of that paper is 31,060 Mflop/s for the NEC SX-4/32. That is a 49% improvement over the results 28 months ago described in the claim above.
