Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!sdd.hp.com!ucsd!nosc!logicon.com!trantor.harris-atd.com!charybdis!sonny From: sonny@charybdis.harris-atd.com (Bob Davis) Newsgroups: comp.binaries.ibm.pc.d Subject: Re: Experience with bargain 80287 Message-ID: <4053@trantor.harris-atd.com> Date: 2 Aug 90 01:10:14 GMT References: <12085@sdcc6.ucsd.edu> <1238400005@uxa.cso.uiuc.edu> <109498@kean.ucs.mun.ca> <1990Aug01.175639.10728@looking.on.ca> Sender: news@trantor.harris-atd.com Reply-To: sonny@trantor.harris-atd.com (Bob Davis) Organization: Advanced Technology Dept., Harris ESS., Melbourne, FL Lines: 48 In article <1990Aug01.175639.10728@looking.on.ca> brad@looking.on.ca (Brad Templeton) writes: [Some stuff deleted] > >Everybody's posting speed diffs between software math and AMD's 287. That's >boring, because if you knew how bad most of the software math packs are, >you would not be impressed by a 5 time speedup. >-- >Brad Templeton, ClariNet Communications Corp. -- Waterloo, Ontario 519/884-7473 Yes. Sometimes the SOFTWARE emulation of FULL coprocessor hardware-precision is VERY slow. Perhaps there is no way around this--I do not know. What I do know is that in Microsoft FORTRAN, doing full-blown coprocessor accuracy using software emulation routines, a certain floating point-intensive benchmark of mine takes: 945 seconds (full coprocessor accuracy FP software routines) Adding the AMD 80C287 coprocessor reduces the time by a factor of 4.7: 202 seconds (80C287 coprocessor, full floating point accuracy). HOWEVER, if one accepts the reduced accuracy available with the Microsoft FORTRAN Alternate Math Library (/FPa option), the benchmark time WITHOUT A COPROCESSOR drops to: 290 seconds (No coprocessor, reduced accuracy FP routines). I have found in several spectral analysis applications making heavy use of floating point-intensive FFT analysis routines, that the accuracy using /FPa is completely adequate. For such applications where one can use the reduced accuracy software floating point routines, one will be much less fascinated with the speedups associated with an expensive coprocessor (in my case, quoted above, I went from 290 seconds with no coprocessor, to 202 with it--both at acceptable floating point accuracies). The bottom line is: use reduced accuracy, hence faster, software routines if you can accept the results. You may find that a coprocessor is more of a luxury than you first thought. However, if money is no major obstacle, go for the coprocessor--there WILL be applications where the full floating point accuracy is required. ______________________________________________________________________________ Bob Davis \\ INTERNET : sonny@trantor.harris-atd.com | _ _ | Harris Corporation, ESS \\ UUCP : ...!uunet!x102a!trantor!sonny |_| |_| | | Advanced Technology Dept.\\ AETHER : K4VNO |==============|_/\/\/\|_| PO Box 37, MS 3A/1912 \\ VOICE : (407) 727-5886 | I SPEAK ONLY | |_| |_| | Melbourne, FL 32902 \\ FAX : (407) 729-2537 | FOR MYSELF. |_________|