Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP Path: utzoo!mnetor!seismo!rutgers!ucla-cs!sdcrdcf!trwrb!desint!geoff From: geoff@desint.UUCP (Geoff Kuenning) Newsgroups: comp.sys.m68k,comp.sys.intel Subject: Re: small instruction cache (was: Recent Motorola ad seen in Byte) Message-ID: <556@desint.UUCP> Date: Thu, 16-Apr-87 14:43:02 EST Article-I.D.: desint.556 Posted: Thu Apr 16 14:43:02 1987 Date-Received: Sun, 19-Apr-87 12:58:05 EST References: <362@sbcs.UUCP> <1466@ncr-sd.SanDiego.NCR.COM> <580@plx.UUCP> <513@omen.UUCP> <285@winchester.mips.UUCP> <518@omen.UUCP> Reply-To: geoff@desint.UUCP (Geoff Kuenning) Organization: Interrupt Technology Corp., Manhattan Beach, CA Lines: 28 Keywords: Sieve, 68020 vs. 80386, Cray, CDC 6600, CDC 7600, cache Summary: the cdc 6600 had a TINY but useful cache Xref: mnetor comp.sys.m68k:365 comp.sys.intel:161 In article <518@omen.UUCP> caf@.UUCP (PUT YOUR NAME HERE) writes: > Does somebody have real data on how the 68020's cache improves performance > on sieve, troff, sort, and other Unix CPU hogs? Well, I don't know about UNIX codes, but data from the old days of supercomputers is quite relevant. The CDC 6600 had what was called an "instruction stack" of about 8 words, or a maximum of 32 instructions. The 7600 expanded this to about 25 words (I'm sure I'm off on the exact numbers), or about 100 instructions. These were not true caches; they were only useful on loops that had backwards branches. The "stacks" worked by saving the last n words of instructions executed. As long as you ran straight-line code, you didn't get any advantage. However, if you branched backwards (looped), the machine would try to find the branch target in the "stack" (cache) and would then run out of cache for the duration of the loop. There is absolutely NO question that these caches improved performance on many useful Fortran applications programs. In fact, the 7600's instruction stack was one of the causes of the downfall of the CDC STAR-100, because it turned out that an in-stack hand-coded vector loop on the 7600 could outrun the Star's special vector instructions. -- Geoff Kuenning geoff@ITcorp.com {hplabs,ihnp4}!trwrb!desint!geoff