kurt@fluke.UUCP (Kurt Guntheroth) (04/30/84)
I have some of the offending Intel Benchmark Report(s). They are:
1 "Benchmark: iAPX 88 versus Zilog Z80", 1980 AFN 01664A
2 "Benchmark: iAPX 88 versus MC6809", 1980 AFN 01532A
3 "iAPX 86 System Benchmark Report", Feb 1982, Order #210352-001
4 "iAPX 186,286 Benchmark Report", Oct 1982, Order #210826
These benchmarks contain some interesting information:
1 and 2. Even on benchmarks optimized for the iAPX 88, the MC6809 is shown as
about 50% of a iAPX 88 (8088). I think this is a pretty good showing for an
8-bit processor. The 88 is only about 3 times as powerful as the lowly Z80
according to these benchmarks. Pretty amazing. Of course, the 88 is
probably much easier to program in assembly language than the Z80.
3. This benchmark compared an Intellec running at 5Mhz against a Exormacs
running at 8 Mhz, with the results normalized for 8 Mhz operation. Note one
important bias in this test was that the Intellec had an 8087 numeric
coprocessor, while fp in the 68000 was interpreted. This alone explains the
good showing of the 8086 against the 68000. Add to this the fact that the
Motorola Pascal used is known not to be a particularly efficient
implementation, even for the 68000, and you see how the results were
obtained. This report makes other comparisons showing the iAPX 86 off
against minicomputers. There is less data about how the tests were
performed, and therefore less discernable bias.
4. The results of these benchmarks are difficult for me to interpret
because they rely on intricate configurations of the hardware. I have heard
rumors that these hardware configurations may not have been completely fair,
especially in the way the MMU was connected to the 68000, and it also
appears that some of the results were obtained by reference to timing charts,
rather than on actual hardware. I will not analyze these results, except to
mention that the 186/286 of course beat the 68000 again. These benchmarks
were mostly on computationally intensive problems like matrix multiply and
FFTs and I did not find any mention of whether a floating point accelerator
was used for either processor.
What conclusions can be drawn from these benchmarks? The 8087 is a
vital piece of any iAPX-x86 based system. The 8088 is not that great a
leap beyond 8-bit processors. An x86+8087 architecture will probably
outperform the 68000 alone in numerical applications. Microcomputers
are achieving performance equivalent to minicomputers.
What conclusions cannot be drawn from these benchmarks? That is for you to
decide. Order these benchmarks and make your own analysis.
Oh, by the way, if you want to see sparks, ask the Motorola rep about their
response to the Intel benchmarks. Just don't stand too close to him when
you do.
Kurt (why doesn't anybody benchmark against the NSC16000 ?) Guntheroth
Not affiliated with Intel, Motorola, NSC, or any other processor manufacturer.
Always my own opinion and doesn't reflect opinions of John Fluke Mfg. Co., Inc.
--
Kurt Guntheroth
John Fluke Mfg. Co., Inc.
{uw-beaver,decvax!microsof,ucbvax!lbl-csam,allegra,ssc-vax}!fluke!kurtdan@idis.UUCP (Dan Strick) (05/01/84)
Pardon me if I have erroneously jumped to the obvious conclusion that the numbers uicsl!keller reported in his usenet news article (uicsl.7000059) mean that Intel is trying to give the impression that typical iAPX 286 systems are roughly twice as fast as typical VAX 11/780 systems for some reasonably large subset of typical applications. I haven't seen the article in the March 1984 issue of IEEE Spectrum which contains these numbers. I have seen the advertisements on page 67 of the February 1983 issue of MINI-MICRO SYSTEMS and page 252 of the June 1983 issue of MINI-MICRO SYSTEMS. I saw stuff like "three times the performance of what you thought was the fastest chip in the market." I saw bar charts that showed the 80286 to be roughly equivalent to three 68000s or one VAX 11/780. The fine print is irrelevant when the average reader doesn't know if the stated system configurations are typical or what the benchmarks really are. The unmistakable implication of the February advertisement was that the 80286 was generally available at that time as was the 68000 and that the typical 80286 system would run a typical cpu bound application about three times as fast as a typical 68000 system. None of this was true then. 80286 systems are just now becoming generally available and actual measurements I made of a 5 MHz 80286 system running with an unknown number of wait states suggest that 8 MHz 80286 systems running with zero wait states will be about as fast as 10 MHz 68000 systems running with zero wait states (which have been available for the last year or so). I have reviewed the article I posted nine days ago (idis.280) which omsvax!plb finds so objectionable and I stand by it. I never said that Intel had fibbed about the benchmarks. I did say that the Intel marketing department had been publishing misleading benchmarks. If Mr. Barrett had looked at my article carefully, he would have noticed that it said absolutely nothing about any conclusions that uicsl!keller may have drawn from the benchmarks. Daniel R. Strick [decvax|mcnc]!idis!dan