HALLAM@vax1.physics.oxford.ac.uk ("Phillip M. Hallam-Baker") (03/04/91)
Dear all, Just a thought - if there are going to be a whole load of even faster computers comming just arround the corner why don't we just shelve all the research into *really* compute intensive stuff (fluid dynamics etc) untill those faster processors arrive. We could spend the time saved drinking beer or something. ...or maybe not! Of course there are some jobs that you just can't do on any single processor that is likely to be built this millenium (aint that a nice pompous phrase ?) - finite element analysis of every nut an bolt in the eifiel tower for example. but then again have a peek at how the latest chips get their massive increases in power - application of parallelism within the chip. Current trends are for pipelined processors with multiple compute units on each chip look at the rapidly increasing disparity between claimed and experienced performance. I dimmly remember from my VLSI lectures that tas you cut down the minimum feature size your chip speeds up - but you also get more area which is equally important. So what do you do with the extra area ? - parallelise go for 5 bit multiply rather than 3 (per minicylce) etc etc. What do you do when you run outa ecconommic accelerators, cache RAM - but that dosen't add to the `headline performance' figure. So yo soon get to the point where you have more than one compute unit. Parallel processing is here to stay. The only question is whether multiple chipping is ! Phill Hallam-Baker Oxford Nuclear Phisics Lab Disclaimer : the usual.. Ps Our application needs parallel processing - we have a 1Tbyte/sec raw data rate - you can't analyse that with a single silicon processor even if you can draw lines at the quantum limit (so there!)