mash@mips.COM (John Mashey) (12/31/89)
In article <789@stat.fsu.edu> mccalpin@stat.fsu.edu (John Mccalpin) writes: >In article <788@stat.fsu.edu> I wrote: ..... >To which brooks@maddog.llnl.gov (Eugene Brooks) replied: ..... >By the way, the most cost-effective machine on my code is the new >Stardent 3000. It runs at about 1/15 of the speed of the Cray on a >per-cpu basis and is less than 1/50 of the cost.... Too bad I can't >afford one! My brain may be jet-lagged, but if I ignore the random shots, I think what I heard was: 1) The protagonists in this argument, both of whom have access to various flavors of supercomputers and Killer Micros, and either of whom would happily consume infinite bunches of cycles: a) Mostly agree on the fundamentals, which I think are: a1) Supercomputers are still way ahead of KMs on raw performance for vectorizable stuff. a2) KMs are ahead of supercomputers on cost-performance for scalar things. a3) Supercomputers are still Good Things, and they'd like them to be free, but since they aren't, it's rapidly becoming a good idea to move scalar codes to KMs when possible, and let the Supers spend more of their time doing what they do best. a4) Vector KMs (like Stardent) might be ahead of supers on cost/performance for at least some vector codes. a5) The KMs (as a group) are gaining, at least somewhat. b) Maybe disagree a little on: b1) Whether KMs are ahead on raw scalar performance (since some of this depends on exactly when the newest KMs ship, and what supers/minisupers appear in the same timeframe. And of course, this all depends on the benchmark... b2) Whether scalar KMs are effectively (i.e., in the necessarily large configurations) ahead in cost/performance on vector probs. b3) Exactly how fast the KMs are gaining. I think this whole argument can interestingly be reduced to filling in this chart [which I can't do, but maybe others can]: Relative Performance of KMs [pick some] versus some supers [pick some] in 1990: KM type Problem Type Scalar Scalar Vector Vector Paral. Paral. Vec-Par Vec-Par Perf Cst/Per Perf Cst/Prf Perf Cst/Prf Perf Cst/Prf Fast Scalar Uni >= > < ? Scalar Multi > Vector Multi > < > So, for instance, pick RC6280, SGI 4D/xxx, Stardent, vs (some) current Cray, and try filling in the boxes with <, =, or >, at least (i.e., < means the KM is less good than super, etc). (I've filled in what I thought I heard, with the ? showing an argument.) Is that at least a more definite framework for the argument? If not, what is? -- -john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc> UUCP: {ames,decwrl,prls,pyramid}!mips!mash OR mash@mips.com DDD: 408-991-0253 or 408-720-1700, x253 USPS: MIPS Computer Systems, 930 E. Arques, Sunnyvale, CA 94086
mccalpin@stat.fsu.edu (John Mccalpin) (12/31/89)
In a short series of articles, Eugene Brooks and I have been flaming back and forth (in a reasonably light-hearted sort of way) about the relative merits of vector supercomputers vs KILLER MICROS. Then along comes John Mashey asking us to be more clear about what we agree and disagree on. In article <34030@mips.mips.COM> mash@mips.COM (John Mashey) writes: >My brain may be jet-lagged, but if I ignore the random shots, I think >what I heard was: > >1) The protagonists in this argument, both of whom have access to various >flavors of supercomputers and Killer Micros, and either of whom would happily >consume infinite bunches of cycles: > a) Mostly agree on the fundamentals, which I think are: > a1) Supercomputers are still way ahead of KMs on raw > performance for vectorizable stuff. > a2) KMs are ahead of supercomputers on cost-performance for > scalar things. > a3) Supercomputers are still Good Things, and they'd like them > to be free, but since they aren't, it's rapidly becoming > a good idea to move scalar codes to KMs when possible, > and let the Supers spend more of their time doing what > they do best. > a4) Vector KMs (like Stardent) might be ahead of supers on > cost/performance for at least some vector codes. > a5) The KMs (as a group) are gaining, at least somewhat. I think that Eugene and I agree on all of these, except that in the last statement, we would both modify it to say that the KM's are gaining ground *very fast*. Another modification of (a4) would be to comment that the mini-supers (Convex) are also typically more cost-effective than top-of-the-line supercomputers on vector codes. > b) Maybe disagree a little on: > b1) Whether KMs are ahead on raw scalar performance (since some > of this depends on exactly when the newest KMs ship, and > what supers/minisupers appear in the same timeframe. And > of course, this all depends on the benchmark... > b2) Whether scalar KMs are effectively (i.e., in the > necessarily large configurations) ahead in cost/performance > on vector probs. > b3) Exactly how fast the KMs are gaining. I think that these are all nicely debatable topics :) There is no question that the KM's have been gaining ground very fast. There are some questions in my mind as to how long this will continue, but the KM's certainly have a strong economic edge in marginal returns. The ETA-10 required >$400 million in its development, which effectively bankrupted CDC. The Cray-3 has already consumed over $300 million in development costs and is still a good 18 months way from being a marketable product. I don't know what MIPS development costs were for the RC6820, but I would bet that they were a bit smaller ! This will continue to act in favor of the KM developers for the next several years, at which point there may not be any supercomputer vendors to compete against. >I think this whole argument can interestingly be reduced to filling in this >chart [which I can't do, but maybe others can]: I'll give it a try! Relative Performance of KMs [pick some] versus some supers [pick some] in 1990: Problem Type --------------------------------------------------------------- | Scalar Scalar Vector Vector Paral. Paral. Vec-Par Vec-Par KM type | Perf Cst/Per Perf Cst/Prf Perf Cst/Prf Perf Cst/Prf ------------- --------------------------------------------------------------- Fast Scalar Uni > >> << <= ? ? <<< <= Scalar Multi > >> << <= ? ? << <= Vector Multi > >> < > ? ? < > >So, for instance, pick RC6280, SGI 4D/xxx, Stardent, vs (some) current Cray, >and try filling in the boxes with <, =, or >, at least (i.e., < means the >KM is less good than super, etc). (I've filled in what I thought I heard, >with the ? showing an argument.) >Is that at least a more definite framework for the argument? If not, what is? >-john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc> Whatever made you think that an argument needed a framework? :-)