Ultracomputers: a teraflop before its time

Computer duihave been 'trivhq for r decade to build rupcrcompum that nm at speeds ncnr one temb (10" h t i n g point operatiom per E C O ~ ~ . Accelerating this achievement would require rho dsvebpmonc of what 1 temr "ultr;lcomputsn" (1) h t heavily rely on parallel pmauirq. Akosdy, htqtncration ulnacompum an available, coosisttng of rhoutands of ntworked cornputen and cordng bctween $50 &n to $300 million (FUp. 1). But d m m a c h yicld high performnce only in rpccialized, highly parallel applicationr, and rht in turn nquircs new a1gorfthmr and roftwim. In my judgment, rubstantidly more wcrful computen will be a d a b l e in l99P"chdt will o&r twaAop prhrmancc for the c a t of ptsoenc supercornputan. Wotk fn propem and dcvelopmene on the h o h n pmmk an era of ucommodity supercompudng." Iktter computers will be available in 1995 if the ovmment chat w o u ~ J o n purchtuing present ultracomputera w e e turned hcead toward training and uofouars to exploit the of rhe next gcncradon. In 1989, =bed the tirrrdon UI high-performance computing in science a d enginccrin and ~pecificolty msnttoned nvf' eta1 par& archirecturef that could deliver terdop power by 1W5, asstuning no a h strainw on price @. My prediction wor that either of two dtcmtives could ochkw thir