Path: utzoo!mnetor!tmsoft!torsqnt!lethe!yunexus!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!cs.utexas.edu!swrinde!zaphod.mps.ohio-state.edu!pacific.mps.ohio-state.edu!linac!midway!ux1.cso.uiuc.edu!uxa.cso.uiuc.edu!msp33327 From: msp33327@uxa.cso.uiuc.edu (Michael S. Pereckas) Newsgroups: comp.arch Subject: special purpose architectures (was Re: Computers for users not programmers) Message-ID: <1991Feb9.015302.5003@ux1.cso.uiuc.edu> Date: 9 Feb 91 01:53:02 GMT References: <13615@lanl.gov> <27AF17B9.72E2@tct.uucp> <5275@mentor.cc.purdue.edu> <27B19A39.321E@tct.uucp> <5437@mentor.cc.purdue.edu> <26337@dime.cs.umass.edu> Sender: news@ux1.cso.uiuc.edu (News) Organization: University of Illinois at Urbana Lines: 42 In <26337@dime.cs.umass.edu> yodaiken@chelm.cs.umass.edu (victor yodaiken) writes: >In article Benjamin Chase writes: >> >>Providing hardware support specifically for these things is unwise. >>The expected performance benefit does not justify the required >>additional complexity of the hardware. Trigonometric and exponential >>functions represent a minute fraction of the average instruction usage >>of general purpose computers. We all understand that you are not >>doing general purpose computing. We don't care. We are very sorry >>that your computer usage represents a minority, and thus has less >>clout when it comes time to design hardware, languages, etc., etc. >> >It might be more interesting to have a discussion about the >diffculty of special purpose architectures and the appropriate >achitectures for scientific computation than to have a discussion about >marketing requirements. Programmable architectures, in which users could >configure the machine for a particular algorithm have implications for >architectures, programming languages and operating systems. I know that >there have been some attempts at such architectures (mostly programmable >pipelines), but would like to >hear about current efforts and/or reasons why such designs would be good or >bad ideas. Sounds like a systolic array. Wasn't someone (Intel?) working on some sort of spiffy programmable systolic array? I recall hearing something a while ago. The idea is appealing, but I'm not certain how wide a range of problems it would be useful for. I also wonder whether compilers could effectivly produce code for a systolic array from dusty deck source code. Efficient auto-vectorizing fortran compilers greatly helped the acceptance of vector supercomputers, after all. One imagines that time sharing might be a problem, also, considering how much state information might have to be saved and restored. -- Michael Pereckas * InterNet: m-pereckas@uiuc.edu * just another student... (CI$: 72311,3246) Jargon Dept.: Decoupled Architecture---sounds like the aftermath of a tornado