husmann@uiuccsb.UUCP (05/01/84)
#R:houxk:-25500:uiuccsb:5600014:000:494 uiuccsb!husmann Apr 30 20:58:00 1984 A person from NEC gave a talk here about their >1 gigaflop machine. The machine had 16 pipeline units. The machine did get a gigaflop if all 16 units were operating, but I doubt if a real program exists which has enough computation in parallel to sustain 16 pipeline units for any length of time. I certainly haven't seen anything to indicate they know more than Cray about supercomputing. Harlan Husmann University of Illinois at Urbana-Champaign UUCP: {ihnp4 | pur-ee}!uiucdcs!husmann
jackson@uiuccsb.UUCP (05/01/84)
#R:houxk:-25500:uiuccsb:5600015:000:1127 uiuccsb!jackson Apr 30 21:20:00 1984 /**** uiuccsb:net.arch / houxk!rdt / 6:21 pm Apr 29, 1984 ****/ several publications have claimed hat the japanese have caught up and in some cases exceeded cray-2 level performance. it seems to me that the only way that this could have happened to their american counterparts is that 1. we took them for granted. 2. we are losing a technology driven race. (GaAs vs. Si.) 3. we are running out of hardware speedup techniques. richard trauben /* ---------- */ You seem to have forgotten one major technique of speeding up programs. That is through the use of program restructurers, the way many vectorizers and vectorizing compilers work. Although Fujitsu's VP200 has a lower peak MEGAFLOP rate than the CRAY-XMP, on the benchmarks that I've heard about the VP200 out performs the CRAY because of the superior vectorizing compiler Fujitsu has. That is, Cray has faster hardware than the Fujitsu, but the Fujitsu has superior software to overcome the slower hardware -- thus a faster machine for the user. Dan Jackson University of Illinois at Urbana-Champaign UUCP: {pur-ee | ihnp4}!uiucdcs!uiuccsb!jackson
ags@pucc-i (Seaman) (05/01/84)
> A person from NEC gave a talk here about their >1 gigaflop machine. The > machine had 16 pipeline units. The machine did get a gigaflop if all 16 > units were operating, but I doubt if a real program exists which has enough > computation in parallel to sustain 16 pipeline units for any length of time. > I certainly haven't seen anything to indicate they know more than Cray about > supercomputing. > > > Harlan Husmann > University of Illinois at Urbana-Champaign I can assure you that there are programs running on our Cyber 205 at Purdue which have "enough computation in parallel" (i.e. long enough vectors) to keep the pipes busy. By the way, "vector length" and "number of pipes" are orthogonal concepts on the 205. Vector lengths can go up to 65,535. All vector instructions (regardless of the vector length) use all available pipes in parallel. We have only two pipes, but if we somehow had 16 we could still keep them all busy for over 4000 clock cycles (=200 microseconds) with a single instruction. -- Dave Seaman ..!pur-ee!pucc-i:ags "Against people who give vent to their loquacity by extraneous bombastic circumlocution."