cca@newton.physics.purdue.edu (Charles C. Allen) (05/03/90)
Where can I get ahold of some benchmarks for the i860 processor? Any system configuration is OK (please mention it), but I'm looking at a brochure for the "Mercury MC860" in particular. They claim 80Mflops for a single processor at 40Mhz. What kind of MFlops are those? Charles Allen Internet: cca@newton.physics.purdue.edu Department of Physics HEPnet: purdnu::allen, fnal::cca Purdue University talknet: 317/494-9776 West Lafayette, IN 47907
pjg@acsu.Buffalo.EDU (Paul Graham) (05/07/90)
cca@newton.physics.purdue.edu (Charles C. Allen) writes: |Where can I get ahold of some benchmarks for the i860 processor? Any |system configuration is OK (please mention it), but I'm looking at a |brochure for the "Mercury MC860" in particular. They claim 80Mflops |for a single processor at 40Mhz. What kind of MFlops are those? as i understand it they simply replay the numbers they got from intel. i believe these have some basis in linpack. however (if it matters) despite announcements to the contrary mercury does *not* have a fortran compiler and the people who return phone calls don't seem to know much about it and the people who know about it don't return calls. i decided to buy something else.
preston@titan.rice.edu (Preston Briggs) (05/07/90)
In article <24706@eerie.acsu.Buffalo.EDU> pjg@acsu.Buffalo.EDU (Paul Graham) writes: >cca@newton.physics.purdue.edu (Charles C. Allen) writes: >| They claim 80Mflops >|for a single processor at 40Mhz. What kind of MFlops are those? > >as i understand it they simply replay the numbers they got from intel. >i believe these have some basis in linpack. The 80 MFlop at 40 MHz number is the peak performance. That is, it's the number you will never surpass. Actual performance will be a lot lower, depending on your application and compiler. The best I've achieved is 48+ MFlops at 33 MHz, for a 400x400 sigle precision matrix multiply. This scales to about 64 MFlops if your memory will keep up. Note also that this was seriously restructured code with the inner loop rewritten in assembly. Note also that it's a factor of 21 faster than you'll get with a straightforward optimizing compiler. -- Preston Briggs looking for the great leap forward preston@titan.rice.edu
mccoy@pixar.UUCP (Daniel McCoy) (05/08/90)
In article <3632@newton.physics.purdue.edu> cca@newton.physics.purdue.edu (Charles C. Allen) writes: >Where can I get ahold of some benchmarks for the i860 processor? Any >system configuration is OK (please mention it), but I'm looking at a >brochure for the "Mercury MC860" in particular. They claim 80Mflops >for a single processor at 40Mhz. What kind of MFlops are those? Peak hand-coded pipelined assembly language. Guaranteed not to exceed. In my experience with porting a heavy scalar floating point renderer (100,000+ lines of C), using a beta release Greenhills compiler from last september, with -O turned on only in the most critical areas (the code wouldn't run otherwise, I did say beta), a 33MHz i860 came in with times faster than an R2000 (DecStation 3100) but slower than a 25Mhz R3000 (SGI-4D/2xx) (MIPS 1.31 compiler with -O2 everywhere). Intel has said that they will release Specmarks soon. Probably around the time that a supported released compiler is available. (Over a year after the announcement of the i860.) I think the Specmarks will inject a much needed dose of reality into Intel's marketing hype. Don't get me wrong, the i860 is a pretty fast chip so the reality isn't that bad. I'm sure a 40Mhz i860 with a well designed memory system will be competitive. But that 80 MFlop number is only meaningful for applications that have really tight main loops that can be hand-coded, and that happen to fit the i860 pipeline model well. The tools have a way to go to catch up with MIPS. The chip above might actually be faster than the R3000, but it doesn't matter to me until a compiler can push it faster. Touting this as a RISC chip (whatever that means) and then saying you should hand code your loops and use a machine level debugger seems a little strange to me. Dan McCoy {ucbvax,sun}!pixar!mccoy (Standard disclaimers apply. Personal opinion only.)
pjg@acsu.Buffalo.EDU (Paul Graham) (05/08/90)
preston@titan.rice.edu (Preston Briggs) writes: |pjg@acsu.Buffalo.EDU (Paul Graham) writes: |>cca@newton.physics.purdue.edu (Charles C. Allen) writes: |>| They claim 80Mflops |>|for a single processor at 40Mhz. What kind of MFlops are those? |> |>as i understand it they simply replay the numbers they got from intel. |>i believe these have some basis in linpack. |The 80 MFlop at 40 MHz number is the peak performance. gahhhh, indeed the linpack value i heard tossed about was 10-13 for what i assume was the 100x100 matrix. 80 is guaranteed . . . not to be exceeded. sorry about any confusion.