Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP Path: utzoo!watmath!clyde!cbosgd!ihnp4!houxm!whuxl!whuxlm!akgua!gatech!seismo!hao!hplabs!oliveb!glacier!mips!mash From: mash@mips.UUCP Newsgroups: net.arch Subject: Re: Addressing modes Message-ID: <399@mips.UUCP> Date: Sat, 15-Mar-86 01:08:26 EST Article-I.D.: mips.399 Posted: Sat Mar 15 01:08:26 1986 Date-Received: Mon, 17-Mar-86 04:41:45 EST References: <946@garfield.UUCP> <1417@sdcsvax.UUCP> Organization: MIPS Computer Systems, Sunnyvale, CA Lines: 58 Greg Noel (Greg@ncr-sd.UUCP or Greg@nosc.ARPA) writes: > In article <389@mips.UUCP> mash@mips.UUCP (John Mashey) writes: > > cycles/MIPS = (clock rate in Mhz) / (MIPS rating) > >so that, for example, if you think a VAX 11/780 with UNIX is a 1 Mips cpu, > >at 5Mhz, it gets a 5. If you think a good 68020 implementation is > >twice as fast, then it gets 16.67/2 = 8.33, etc. [Notice that this > >isn't cycles/instruction, but cycles/equivalent work.] > > I'm currious -- are these off-the-top-of-the-head numbers to illustrate the > point, or do they approximately reflect reality? If the latter, are there > any published results? 1) They're not off-the-top-of-the-head numbers, but do reflect reality, given enough caveats and specifications, always remembering that: a) Reducing CPU performance to 1 number is always hard. This doesn't ever stop anybody in this buisness from trying. b) One must always include software: thus, a more complete description of the above might be: ".. if you think a VAX 11/780 with 4.2BSD runs user programs like yacc, grep, diff, etc..., compiled with the regular C compiler, and runs them at at rate we'll call 1 Mips, at 5 Mhz, it gets a 5. If you think a 68020 (SUN 3/160, running their currently-released software) is twice as fast.... it gets 8.33." 2) SUN has published a bunch of comparisons as part of their SUN-3 announcement. I think Convergent has published theirs for the MightyFrame, and Altos for their 3068, and ..... 3) I personally believe these numbers are probably good within +/- 20%p what I see personally, when I look [which is not real often or very carefully), is that the SUN is 2X or a little more faster for some things, especially little benchmarks, and is about 1.8X for many tasks. 4) As I've observed before here, IT IS VERY IMPORTANT TO COMPARE ACTUAL MACHINE PERFORMANCE AND NOT GET CONFUSED BY MIPS RATINGS. In particular, many people do call a 780 a 1Mips (or thereabouts) machine, but it is known that for some fairly substantial runs (user and kernel), the 780 averages 10.6 (200ns) cycles / instruction, i.e., it does .5M VAX instructions/sec, including all of the degradations for memory, TLB misses, pipeline bubbles for branches, etc, etc. [There was a good 1984 paper by Clark and Emer of DEC on this.] Thus, it would appear that a machine that does 1 Million VAX instructions/sec (i.e., a bit faster than a 785) does an amount of work like what people usually call a 2Mips machine. Thus, a machine that executes 1M of its own instructions/sec to do the same work as the 780's .5M VAX instructions can also be called a 1Mips machine, since people call th4e VAX a 1Mips cpu. Thus, there would seem to be twice as many, but "twice as simple" (whatever that means) instructions. So, in some sense, the instruction within the mythical "1Mips" is probably more like a RISC instruction than a VAX (CISC) instruction! Again, what counts is having a scale that people understand and on which systems can be compared consistently. [At MIPS, we just always call the given flavor of 780 == 1 and be done with it.] -- -john mashey UUCP: {decvax,ucbvax,ihnp4}!decwrl!mips!mash DDD: 408-720-1700 USPS: MIPS Computer Systems, 930 E. Arques, Sunnyvale, CA 94086