This is interesting news from NASA’s Ames Research Center:
In collaboration with physicist James Freericks of Georgetown University, Sherry Chang and Johnny Chang of the NASA Ames Supercomputing (NAS) Division recently co-authored a paper on some impressive optimization and scaling work done on Freericks’ research code, NDMFTS. The paper describes NAS’ efforts to optimize and scale NDMFTS, ultimately achieving a 7.3-fold speedup and 8.53 teraflops (66% of peak) sustained performance on 2,032 processors—the highest sustained teraflop performance of any application run on Columbia.
Recall that Columbia is NASA’s Itanium-based SGI Altix 3700 (specs here).
Of course we already know that if we work hard enough we can often get good performance. What we want from our next machines is to get good performance without working so very hard, and without targeting so much of that work to one specific machine.