home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!cs.utexas.edu!sun-barr!news2me.ebay.sun.com!cronkite.Central.Sun.COM!sixgun.East.Sun.COM!chessie!pmorris
- From: pmorris@chessie.East.Sun.COM (Phillip Morris - SE Washington D.C.)
- Newsgroups: comp.arch
- Subject: Re: Call for Opinion: Viking or i860
- Date: 31 Jul 1992 11:57:38 GMT
- Organization: Sun Microsystems, Inc.
- Lines: 58
- Distribution: world
- Message-ID: <15b9riINN3o4@sixgun.East.Sun.COM>
- References: <5678@nosc.NOSC.MIL>
- Reply-To: pmorris@chessie.East.Sun.COM
- NNTP-Posting-Host: chessie.east.sun.com
-
- In article 5678@nosc.NOSC.MIL, wolfgang@sunspot.nosc.mil (Lewis E. Wolfgang) writes:
- >Call for opinion:
- >
- > We run large acoustic modeling programs on a
- >variety of Sun Sparc systems and in the past have
- >considered purchasing array processors to hasten things
- >along. It seems as if many array processors are based
- >on Intel's i660 chip, certainly a respectable
- >numerical masticator, but how does it stack up against
- >current RISC based CPUs?
- >
- > We have a Sun 690MP that we could upgrade to
- >SuperSparc (Viking) with SuperCache for not much more
- >money than an equal number of i860 based processors.
- >It seems as if integer performance is better with Sparc,
- >with floating point not quite as good. An advantage of
- >a Sparc upgrade is that old code would not have to be
- >modified, at a cost of slightly less relative floating
- >point performance.
- >
- > What do you think? Upgrade the cpu with Viking?
- >Or purchase an equal number of i860 based array processors?
- >
- > Thanks,
- > Lew Wolfgang
- > wolfgang@nosc.mil
- >
-
-
- Mr. Wolfgang,
-
- After noting that I work for Sun, please also be aware that this is a new position (only been here
- 3 months) and that my previous jobs were at ATT Bell Labs & BBN, both positions as a near-real-time
- programmer on Sun SparcStations. Most of the stuff we did was digital signal processing which is
- quite floating point intensive. For a long while we used either TI or ATT DSP chips, but when our
- computing needs got more general than just the dsp code, we needed a more general floating point
- processor. We tried out both the Sky & Mercury i860 boards (4 processors) and found that the only
- way to get anywhere near the stated MFLOPS rating was by hand coding in assembly. The C
- cross-compilers stank!! Any 1st year CS student could have optimized better for a particular cpu
- than those compilers did.
-
- All this is to say this: Don't pay any attention *AT ALL* to stated performance figures. Either
- get a loaner piece of equipment and try out your code, or package the demanding part of your code
- into a easily compilable/runnable package and send it to the various manufacturers to let them
- perform benchmarking (and make SURE to get exactly how they compiled and ran it). Then you have a
- basis for fair comparison.
-
- My guess, based upon previous experience, is that unless you plan to hand code assembly for the i860,
- then you will get *better* FP performance out of the SuperSparc (Viking) with the new SparcCompiler
- that has been optimized for that architecture.
-
-
- ---
- ------------------------------------------------------------------------------
- -Phil Morris, SE w/ Sun Microsystems in Vienna, VA
- -pmorris@chessie.East.Sun.COM
-
-
-