mash@mips.UUCP (John Mashey) (10/01/86)
In article <200@mipos3.UUCP> kds@mipos3.UUCP (Ken Shoemaker ~) writes: >I can't speak for Intel, and none of this should be taken as the opinion >of Intel, but... >...Motorola's "announcement" came in the same week that they chided >Intel for the 386's non-availability; the week that Compaq introduced >their box. > >To put things in a historical sense, Motorola held the same kind of blitz >a few months after Intel came out with the 286... >.... I suppose that Intel should feel flattered in that it seems >to me that Mot, in each case, has conceeded defeat: the 68010 to the >286, the 68020 to the 386, and that in order to maintain any kind of >market share, they have to have a media blitz, rolling out their newest, >half baked, family jewels. Its all pretty comic! > >And what about the new family jewels? Should I feel vindicated that Mot >has finally decided that on-chip memory management is a good idea? And >what about performance? Why would a new, ultra-high end user choose >the 68030 over, say, the Clipper, or the MIPS chips, each of which >(if you believe the numbers) should exceed the 68030's performance (with >the promise of delivering more sooner) and are available, in some semblance, >today? Normally, I to stay away of Intel vs Motorola wars (it's more fun watching the missiles go overhead!), but since MIPS' name got dragged in, and since there have been an awful lot of fuzzy performance claims thrown around lately in this newsgroup, and in advertising, I couldn't resist. Several recent Intel and Moto ads have cited Dhrystones to prove performance: 1) The Intel ad claimed that the 386 was faster than any other 32-bit micro, and was a 4-VAX-MIP machine because it did 6133 Dhrystones at 16Mhz, versus a VAX 11/780's 1662. 2) Then, a Moto ad appeared that complained about the unreality of the Intel hardware environment used to get those numbers, versus reality of using a 25Mhz 68020 in a SUN-3/200. (This was the apples-to-oranges ad that appeared lots of places). Some of this was pretty reasonable, but they did fudge a little, claiming that Dhrystone writes had to go to main-memory DRAM. The SUN-3/2xx boxes use a big 64K write-back cache, with a well-designed memory subsystem, i.e., a pedal-to-the-metal, all-out design. This got 6362 Dhrystones. 3) Now: *REALITY*: a) Dhrystones aren't all that good indicator of overall performance. An 8600 gets about 5,000-6,000 Dhrystones, and people (accurately) characterizes an 8600 as about 4-4.2X an 11/780 performance. The benchmark randomly happens to correlate with reality, somewhat. b) Small, individual benchmarks just don't mean a whole lot. Dhrystone can chew up to 30% of the total time in strcpy(). We certainly don't use it for serious performance characterization. c) People who talk about "this chip is an N-Mips chip" are missing the boat: you have to say "This chip, at XMhz, with specified memory system, and specified software environment, runs at some speed relative to other machines." There have been some wild commetns in this newsgroup and some other ads that are nearly meaningless. d) Ignore "peak mips" claims that you see in ads: I've recently seen that a 25Mhz 68020 averages 5Mips and does 12.5Mips peak. Likewise, a 16.67Mhz 68020 (as in a SUN-3/1xx) is 8.3Mips peak. An 8Mhz MIPS R2000 does 8Mips peak, but is typically 2-3 times faster doing real stuff than the 16Mhz 68020. Peak rates in non-real environments have little to do with reality. c) (to Moto): I'm not sure SUN-3/2xx's are shipping yet, but they are at least real machines. d) (to Intel): although Compaq 386's are shipping (fair), it will be interesting to see if their memory system lets them equal 8600s, as seems to be claimed. 4) Finally, like I said, I don't believe in Dhrystones, but since I & M keep using it to show how fast they are, in self-defense, see: 11/780 1632 (in-house 4.3BSD) 68020 16.6Mhz 3243 (registers, -O, in-house SUN-3/160, 4.2BSD) 80386 16.6Mhz 6133 (from an Intel ad in last month) 68020 25Mhz 6362 (from a Moto ad, citing SUN 3/200, with 64K writeback cache). 8600 5132-6423 (ULTRIX or 4.3BSD) Clipper 33Mhz ???? (This is claimed as 5Mips, but I haven't seen any numbers published. Can somebody supply them?) R2000 8Mhz 10350 (MIPS M/500 development system, 16K I- + 8K D-cache, 4.3BSD, release 1.0, 8MB memory, 4-way interleaved, up-to-rev produciton, boards. We use them and we're shipping systems to customers; i.e., REAL, if not yet in huge quantities.) WE CALL THIS AN HONEST 5X 11/780 SYSTEM FOR REAL STUFF, NOT 6.5X. R2000 8Mhz 12300 As above, but with global optimization. [I can't tell what versions I & M are citing.] This part was also designed to run at 12.5Mhz and 16.67Mhz, and you can draw your own conclusions on the performance. [No: they don't scale exactly, but Dhrystone is so cache-resident that they come close]. Bottom line: what counts is real benchmarks on real machines; it is difficult to characterize machines without a lot of hard work; specifying a machine by a single nebulous "Mips-rating" (except within same-architecture/ same-software families, perhaps) doesn't work very well. Finally, if anybody believes that the Dhrystone numbers above mean that, for example, that 16Mhz 386s and 25Mhz 68020s are REALLY generally as fast as VAX 8600s, or that an 8Mhz R2000 is close to a an IBM 3083, then....... I own this great bridge back East that I could let go at a good price!.... -- -john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc> UUCP: {decvax,ucbvax,ihnp4}!decwrl!mips!mash, DDD: 408-720-1700, x253 USPS: MIPS Computer Systems, 930 E. Arques, Sunnyvale, CA 94086