home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.sys.sun.hardware:3547 alt.sys.sun:3002
- Newsgroups: comp.sys.sun.hardware,alt.sys.sun
- Path: sparky!uunet!usc!sdd.hp.com!mips!odin!fido!zola!zuni!anchor!olson
- From: olson@anchor.esd.sgi.com (Dave Olson)
- Subject: Re: Sparc-10 performance oddity
- Message-ID: <novj88o@zuni.esd.sgi.com>
- Sender: news@zuni.esd.sgi.com (Net News)
- Organization: Silicon Graphics, Inc. Mountain View, CA
- References: <Brw9Dw.Bn@helios.physics.utoronto.ca>
- Date: Sun, 26 Jul 92 05:21:20 GMT
- Lines: 42
-
- In <Brw9Dw.Bn@helios.physics.utoronto.ca> sysmark@helios.physics.utoronto.ca (Mark Bartelt) writes:
-
- | I've noticed a peculiarity (described below) with the Sparc-10/41. Does
- | anyone have a clue as to what's going on? To compare the Sparc-10 with
- | older Sparcs, I ran an admittedly crude zeroth-order test of non-FP CPU
- | horsepower, namely how much CPU time does it take for "dc" to calculate
- | sqrt(2) N times to D decimal places, for various combinations of N and D.
- | The test was run three times on a Sparc-1+, a Sparc-2, and a Sparc-10/41.
- | (At least I'm *told* it was a 10/41; it was in the booth at a trade show
- | earlier this summer.) Here are the performance ratios:
- |
- | RelativePerf
- | Prec Loops 2:1+ 41:2
- | ---- ----- ---- ----
- | 5 50000 1.72 1.64
- | 20 7000 1.73 1.35
- | 50 1400 1.72 1.24
- | 99 400 1.74 1.20
- |
- | Note that the performance of the Sparc-2 relative to the Sparc-1+ is
- | fairly constant, independent of which test was run. But for the 10/41,
- | the performance (relative to the Sparc-2) differed, depending on which
- | test was run. (Also, running the *same* test repeatedly gave a wider
- | variance of CPU times than one sees on the older Sparcs.)
- |
- | So, a couple questions: First, even in the best case above, the 10/41
- | doesn't perform anywhere near as well as one might hope, based on the
- | published MIPS/SPECint figures. Is this expected?
-
- I seem to remember that dc is somewhat unusual in its access
- patterns because of the way the multiple precision math is
- implemented (I'll have to look at it some more with pixie and
- other tools on our systems).
-
- Indigo r4k vs Indigo isn't as impressive as the spec differences
- either, giving a ratio of about 1.73 (although it doesn't show
- as much variation with precision, and that isn't consistent), while
- the integer spec ratios are closer to 2.5:1.
- --
- Let no one tell me that silence gives consent, | Dave Olson
- because whoever is silent dissents. | Silicon Graphics, Inc.
- Maria Isabel Barreno | olson@sgi.com
-