news@omepd (News Account) (03/12/88)
-------- From: mcg@iwarpo3.intel.com (Steve McGeady) Path: iwarpo3!mcg In article <1806@obiwan.mips.COM> mark@mips.COM (Mark G. Johnson) writes: >Second, a question. The other DARPA core-MIPS paper at the ISSCC >(a 200-MIPS GaAs bipolar device from Texas Instruments) devoted a >segment of the oral presentation to chip yield. They were quite >pleased to reveal their exact percentage yields to date (on this >DARPA-funded project) and to give their yield projections for the >next 12 months or so. > >Could somebody from GE tell us what the yield is on the GE DARPA >core-MIPS chip? TI's data included (a) # of core-MIPS chips built >to date; (b) # of them that are fully-functional, (c) trendline >predictions of (b)/(a) for the near future. This opens a very interesting can of worms. I would like to ask MIPSCo what *their* chip yield is. And lest I be accused of pursuing a hidden agenda, here it is: This newsgroup (comp.arch) tends to focus on a very narrow band of issues concerning the success (intellectually and commercially) of architectures and their implementations. The predominant issues discussed here are performance, "capability", and architectural elegance. All too seldom do we discuss issues that, ultimately are more important to the success of an architecture: -> bug-free, working silicon -> yield (affects price, availability) -> system integration issues -> HW interface complexity -> availability of compatible interface chips -> support tools -> availability of compilers, tools -> development environment -> software support -> debug capabilities (ICE, SW debug, others) -> hardware support -> demonstration designs We've heard much (too much, some would say) from MIPSCo regarding the raw performance of their processor. I, for one, would be interested in hearing some other questions answered, for instance: 1) Who manufactures your silicon, on what process, what yield do you get, and how does/will this influence chip prices? [I suppose this begs the question of whether MIPSCo is primarily a chip, board, or system vendor - I have heard all three as the answer at various points in the last few years]. 2) Without a captive silicon manufacturing establishment, how can your silicon-foundry-provided 2-micron CMOS technology effectively compete with sub-micron technology from manufacturers with captive silicon development technology? 3) What is the availability of development tools (compilers, assemblers, debuggers) on VAXen, SUN's, and PC's? 4) What is the complexity of integrating a MIPSCo chip set into a system? What amount and kind of support HW is needed? As a more general question to this group, what does it take to make an architecture successful? It can't be elegance of design, for (e.g.) the 80386 and the MIPSco processor are each somewhat inelegant in their own ways (for those who don't wish to fill in the blanks, segmentation and register model in the former case, exposure of pipeline and other implementation details in the latter); It can't be raw speed, for slower processors (e.g. the 68020 and the 80386) are still consistently out-selling faster ones; It can't be ease of UNIX ports, because the original 68000 UNIX ports (e.g. Sun-1 and Apollo) were exercises in frustration (usually due to memory management limitations); And it certainly can't be volume of posting in comp.arch, or MIPSco would have gone public long since. My stab at an answer to this question: The answer is that there are many things that contribute, including marketing (both hype and non-hype), but (and forgive me if this sounds circular), what makes an architecture is VOLUME. An architecture is successful if you sell a lot of the chip implementing that architecture. This is because volume lowers the price of the part, encourages third parties to develop software for it, and in general increases familiarity with the architecture, lowering the learning curve for new designs and the "fear factor" many have. Put another way, I would much rather have designed a processor that found its way into every anti-skid braking system, microwave oven and laser-printer than to have designed a processor that was used only in a particular $100,000 software engineering workstation. I've rambled on long enough. Now, to Mr. Johnson's answer about MIPSCo yield - I'm very interested, and I thank him for giving me an excuse to raise the issue. S. McGeady Intel Corp. p.s. - If you counter with 'what is Intel's yield on (e.g.) the 386', I will have to answer that this is considered one of Intel's most proprietary pieces of information, and I couldn't reveal it even if I knew the answer, which I don't. I can say, however, that we shipped 1,000,000 386's in 1987, and expect to ship more than that in 1988, ugly architecture and all.
jimv@radix (Jim Valerio) (03/12/88)
Keywords: In article <2904@omepd> mcg@iwarpo3.UUCP (Steve McGeady) asks: >what does it take to make an architecture successful? The observation that's always been quoted to me, and seems to have some basis in fact, is: Good computers don't sell. Of course, there are so many bad computers out there that this observation could be an artifact of the random probability of any computer succeeding. :-) Less cynically, it seems pretty clear to me that name brands consistently win out over fads. There's something about sizable investments that makes many people turn conservative. -- Jim Valerio jimv%radix@omepd.intel.com, {verdix,omepd}!radix!jimv
marc@oahu.cs.ucla.edu (Marc Tremblay) (03/15/88)
In article <2904@omepd> mcg@iwarpo3.UUCP (Steve McGeady) writes: > >We've heard much (too much, some would say) from MIPSCo regarding the >raw performance of their processor. I, for one, would be interested in >hearing some other questions answered, for instance: > > 1) Who manufactures your silicon, on what process, what yield > do you get, and how does/will this influence chip prices? As far as I know MIPSCo has agreements with three companies to manufacture and sell their processor. Each of the companies has rights to both the 12.5 MHz and 16.7 MHz version. The three companies are: i) LSI Logic Corp ii) Integrated Device Technology Inc. iii) Performance Semiconductor Corp. > 2) Without a captive silicon manufacturing establishment, how > can your silicon-foundry-provided 2-micron CMOS technology > effectively compete with sub-micron technology from > manufacturers with captive silicon development technology? Performance Semiconductor Corp. has a high-speed CMOS process using submicron technology. > 4) What is the complexity of integrating a MIPSCo chip set > into a system? What amount and kind of support HW is needed? I would also like to know more about that one. (MIPS guys?) Marc Tremblay marc@CS.UCLA.EDU ...!(ihnp4,ucbvax)!ucla-cs!marc Computer Science Department, UCLA
hansen@mips.COM (Craig Hansen) (03/16/88)
In article <10355@shemp.CS.UCLA.EDU>, marc@oahu.cs.ucla.edu (Marc Tremblay) writes: > In article <2904@omepd> mcg@iwarpo3.UUCP (Steve McGeady) writes: > >We've heard much (too much, some would say) from MIPSCo regarding the > >raw performance of their processor. I, for one, would be interested in > >hearing some other questions answered, for instance: > > 1) Who manufactures your silicon, on what process, what yield > > do you get, and how does/will this influence chip prices? > As far as I know MIPSCo has agreements with three companies to > manufacture and sell their processor. Each of the companies has > rights to both the 12.5 MHz and 16.7 MHz version. > The three companies are: > i) LSI Logic Corp > ii) Integrated Device Technology Inc. > iii) Performance Semiconductor Corp. These companies also have rights to future, higher-performance designs. > > 2) Without a captive silicon manufacturing establishment, how > > can your silicon-foundry-provided 2-micron CMOS technology > > effectively compete with sub-micron technology from > > manufacturers with captive silicon development technology? > Performance Semiconductor Corp. has a high-speed CMOS process > using submicron technology. All three companies have competitive CMOS technology. > > 4) What is the complexity of integrating a MIPSCo chip set > > into a system? What amount and kind of support HW is needed? > I would also like to know more about that one. (MIPS guys?) For all the belly-aching about multiplexed busses and multiphase clocks, integrating a MIPS chip set is easier than most microprocessors, particularly when you consider that a _high-performance_ micro needs to have external cache hand-crafted by the system designer, whereas on the MIPS chip, the cache is formed from some latches and buffers and off-the-shelf, standard static RAM's, available in quantity from multiple vendors. All the customised hardware to use the SRAMs as caches are on the chip. [The SPARC chip set also requires a hand-crafted MMU design; we have the MMU on chip.] The support HW required is: 1) a 2x frequency oscillator 2) a tapped delay line 3) some '373 latches 4) some 1804 buffers 5) some fast SRAM The number of latches, buffers, and SRAM depends on the size of the caches you choose; anything from 4kb to 64kb for each of the instruction and data caches is permitted in the current parts. This can all be hooked up in cookbook fashion; we provide design data that shows how it all goes together. Because of the processor's high speed, and its write-through caches, writes to memory occur rather frequently, so we also provide a set of gate arrays that form a 4-deep FIFO between the fast processor and a slower main memory system. The use of these gate arrays is optional, but they're easy to use and improve performance about 10% over having a one-deep write buffer. -- Craig Hansen Manager, Architecture Development MIPS Computer Systems, Inc. ...{ames,decwrl,prls}!mips!hansen or hansen@mips.com 408-991-0234
sedwards@esunix.UUCP (Scott Edwards) (03/16/88)
From article <2904@omepd>, by news@omepd (News Account): > -------- > From: mcg@iwarpo3.intel.com (Steve McGeady) > Path: iwarpo3!mcg > > This newsgroup (comp.arch) tends to focus on a very narrow band > of issues concerning the success (intellectually and commercially) > of architectures and their implementations. The predominant > issues discussed here are performance, "capability", and > architectural elegance. All too seldom do we discuss issues > that, ultimately are more important to the success of an > architecture: > [ alot of valid points deleted ] > > S. McGeady > Intel Corp. > IMHO this newsgroup (comp.arch) tends to focus on a very narrow band of RISCy business. I agree with Steve, there is more to computer architecture than raw cpu mips (vips, vups, ticks, clicks, clocks, whatever:-). I think RISC's are great for specific applications where the instruction set is optimized for that specific application and someone is willing to pay the extra $$$$$ for more, fast memory and optimizing compilers. What about architectures designed for software reliability like SWARD and iapx432 (Yes, I know it was slow, but perhaps that was just the implimentation). Is anyone still working in this direction? Or that dataflow stuff (didn't NEC have some sort of graphics processor) is it dead and buried? Or is *EVERYBODY* (except me) convinced that RISC is the wave of the future? Less is more, yesterday's technology today! ;-) -- Scott ***************************************************************************** * My opinions are not shared by my * On the first part of the journey, * * employer or anyone else with half * I was looking at all the life... * * a brain (or even a whole one). * - America * *****************************************************************************
mch@computing-maths.cardiff.ac.uk (Major Kano) (03/17/88)
In article <2904@omepd> mcg@iwarpo3.UUCP (Steve McGeady) writes: > >This opens a very interesting can of worms. I would like to ask MIPSCo ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ (* Absolutely *) (* stuff deleted *) >And lest I be accused of pursuing a hidden agenda, here it is: > > This newsgroup (comp.arch) tends to focus on a very narrow band > of issues concerning the success (intellectually and commercially) > of architectures and their implementations. The predominant > issues discussed here are performance, "capability", and > architectural elegance. All too seldom do we discuss issues > that, ultimately are more important to the success of an > architecture: > > -> bug-free, working silicon > -> yield (affects price, availability) (* And how. On of the first nieveties (forgive my spelling) that I have had to unlearn at College was that fantastic architectures may not be practical in the real world, given the avaliable process technologies. *) > -> system integration issues > -> HW interface complexity > -> availability of compatible interface chips > -> support tools > -> availability of compilers, tools > -> development environment > -> software support > -> debug capabilities (ICE, SW debug, others) > -> hardware support > -> demonstration designs (* Interesting; good points *) (* Stuff deleted *) >It can't be elegance of design, for (e.g.) the 80386 and the MIPSco processor >are each somewhat inelegant in their own ways (for those who don't wish to >fill in the blanks, segmentation and register model in the former case, ^^^^^^^^^^^^ ^^^^^ ** WHAT THE $@#@++%%& HELL ?!? ** Wow ! And I thought Ray Coles (writes for Practical Computing, a UK Magazine) had it in for Intel ! Agreeing, as I do, that the register model doesn't have enough of them, and that even the '386 isn't regular enough, I though that the feature of the '386 that made it so TECHNICALLY advanced (IBM compatibility not withstanding :-) ** WAS ** its memory management and protection model. The 32-bit within-segment addresses are what people have been waiting for for ages. I would question the fact that only 16 bit selectors are avaliable, but I defy anyone to come up, in the near or intermediate future, with an Intel-style memory model that is better than Intel's, without opening up a whole can of voracious memory-eating killer-worms at the descriptor table level. If you don't know what I mean by that, pretend for a moment, that YOU were the person who had to come up with the byte/page granularity kludge in order to make 4GB segments fit in a DT entry. (Maybe that person (or team) could comment themselves, if they use the net). What does everyone think ? As a computer architecture junkie, I would be very interested; as the Intel segments seem (to me) to come in for a lot of flack every so often; I just didn't expect it to come from an Intel employee . *) (* Stuff deleted *) >than to have designed a processor that was used only in a particular >$100,000 software engineering workstation. (* WOW ! That is the sort of processor I personally WOULD like to work on once I've had enough experience; although I would also like to see such a CPU/chip set sold everywhere, of course. *) > >I've rambled on long enough. Now, to Mr. Johnson's answer about MIPSCo ^^^^^^^^^^^^^^^^^^^^^^^^^^^ (* So have I. Hope someone comments on all this. Bye ! *) (* Stuff deleted *) -mch PS: I would be grateful to anyone who POSTS replies or flames or whatever about this if they would also e-mail me, as I go home for the Easter Vacation in two day's time, and our system manager does expire articles every so often ! Thanks in advance, -mch -- Martin C. Howe, University College Cardiff | "You actually program in 'C' mch@vax1.computing-maths.cardiff.ac.uk. | WITHOUT regular eye-tests ?!" -------------------------------------------+-----+------------------------------ My cats know more about UCC's opinions than I do.| MOSH! In the name of ANTHRAX!
wcs@ho95e.ATT.COM (Bill.Stewart.<ho95c>) (03/29/88)
In article <751@esunix.UUCP> sedwards@esunix.UUCP (Scott Edwards) writes:
:RISC's are great for specific applications where the instruction set is
:optimized for that specific application and someone is willing to pay the
:extra $$$$$ for more, fast memory and optimizing compilers. What about
:architectures designed for software reliability like SWARD and iapx432
You've got it backwards - the theory of CISC is that you really need
binary-coded-decimal-graphics-interpreters because your COBOL programs
draw a lot of pie charts, and need them in hardware to go fast.
RISC says make the hardware do the few simple things *everybody* needs,
and do them elegantly and real fast. That way it's easy to write
good compilers, and fast enough to do your custom work in software.
The main arguments between CISC and RISC people are whether simple and
elegant is enough faster to justify the lack of complex features, and
whether it really *is* easier to write compilers good enough that you
don't need hand-coded assembler. RISC architectures are probably more
reliable, since there's less complexity.
You *always* need optimizing compilers - compare Microsoft C 5.0 with the
earliest 8086 compilers - it's taken a long time to produce decent CISC code.
As far as memory speed is concerned, note that *data* memory is
accessed at about the same rate, because you're doing the same work,
and most calculations are to registers rather than RAM.
Code memory speed has to be higher on RISCs, since you generate more
code, but that's what instruction caches are for.
You also mentioned graphics processors and the like. Certainly when
you have very specialized applications that you use a lot, it's worth
building custom hardware. AT&T's DSP-32 digital signal processor does
8 MFLOPS of add and multiply, plus I/O, because speech processing needs
fast crunching; it can get away with minimal interrupt-handling and
small address spaces because it's not trying to be a general-purpose
chip. Similarly, several of the commercial graphics chips do BITBLTs
real fast, but typically use an 80186 or 68000 to deal with communications.
But for general applications, like hacking integers and characters,
RISC is usually better.
--
# Thanks;
# Bill Stewart, AT&T Bell Labs 2G218, Holmdel NJ 1-201-949-0705 ihnp4!ho95c!wcs
# So we got out our parsers and debuggers and lexical analyzers and various
# implements of destruction and went off to clean up the tty driver...
davidsen@steinmetz.steinmetz.ge.com (William E. Davidsen Jr) (03/30/88)
In article <2089@ho95e.ATT.COM> wcs@ho95e.UUCP (46323-Bill.Stewart.<ho95c>,2G218,x0705,) writes: | You've got it backwards - the theory of CISC is that you really need | binary-coded-decimal-graphics-interpreters because your COBOL programs | draw a lot of pie charts, and need them in hardware to go fast. | RISC says make the hardware do the few simple things *everybody* needs, | and do them elegantly and real fast. That way it's easy to write | good compilers, and fast enough to do your custom work in software. Actually a CISC processor is in reality two parts: a RISC processor which executes a simple native language, and a silicon compiler which takes a pseudo assembler (what we see as the native instruction set) and translates to the native RISC instructions (microcode). Seriously, when I see people claim that they are taking a pseudocode generated by <your favorite compiler> and doing a translate and optimize so they can run it on RISC, I wonder why all the fuss? In a few more years, perhaps ten, we will know how to build a translator in silicon which does a better and faster job of translation, NOP fills, etc, than we are now doing in software. I think we will still do a better job with software at that time, but it may not be cost effective. IBM sort of used this approach with the PC/370, modifying a 68000 to translate 370 opcodes into the same microcode that the 68000 conventional opcodes use. -- bill davidsen (wedu@ge-crd.arpa) {uunet | philabs | seismo}!steinmetz!crdos1!davidsen "Stupidity, like virtue, is its own reward" -me
jesup@pawl14.pawl.rpi.edu (Randell E. Jesup) (03/31/88)
In article <10170@steinmetz.steinmetz.ge.com> davidsen@crdos1.UUCP (bill davidsen) writes: >Actually a CISC processor is in reality two parts: a RISC processor >which executes a simple native language, and a silicon compiler which >takes a pseudo assembler (what we see as the native instruction set) and >translates to the native RISC instructions (microcode). Well, I wouldn't put it that way. More like a silicon interpreter than a compiler (it takes a higher-level language one instruction at a time, interprets what it means, does it, repeat. A compiler would read it it, then write out the microcode version of it for later execution. >Seriously, when I see people claim that they are taking a pseudocode >generated by <your favorite compiler> and doing a translate and optimize >so they can run it on RISC, I wonder why all the fuss? In a few more >years, perhaps ten, we will know how to build a translator in silicon >which does a better and faster job of translation, NOP fills, etc, than >we are now doing in software. I think we will still do a better job with >software at that time, but it may not be cost effective. CISC vs RISC is all a matter of where the balance lies. Neither is 'better', each is better for specific cases. The balance lies in the application it's designed for, the process used for the chip, the speed and size and layout of external memory, cost considerations, etc, etc. In ten years we will have smarter and better CISCs, and faster RISCs. The balance will shift back and forth, but I doubt either will be eliminated. >IBM sort of used this approach with the PC/370, modifying a 68000 to >translate 370 opcodes into the same microcode that the 68000 >conventional opcodes use. Actually, I'd thought they re-microcoded the 68000's to understand 370 code instead of 68000 code (running a different internal interpreter program, effectively.) No translation is done. I think one of the reasons RISC became popular is that on-chip references (of microcode) are no longer immensely faster than off-chip, due to things like pipelining, and chip tech allows denser, faster chips and rams. On another point: Does anyone at Mips want to comment on the R-3000? (Announced a few days ago, I believe) // Randell Jesup Lunge Software Development // Dedicated Amiga Programmer 13 Frear Ave, Troy, NY 12180 \\// beowulf!lunge!jesup@steinmetz.UUCP (518) 272-2942 \/ (uunet!steinmetz!beowulf!lunge!jesup) BIX: rjesup (-: The Few, The Proud, The Architects of the RPM40 40MIPS CMOS Micro :-)
david@titan.rice.edu (David Callahan) (03/31/88)
In article <10170@steinmetz.steinmetz.ge.com> davidsen@crdos1.UUCP (bill davidsen) writes: >Seriously, when I see people claim that they are taking a pseudocode >generated by <your favorite compiler> and doing a translate and optimize >so they can run it on RISC, I wonder why all the fuss? In a few more >years, perhaps ten, we will know how to build a translator in silicon >which does a better and faster job of translation, NOP fills, etc, than >we are now doing in software. I think we will still do a better job with >software at that time, but it may not be cost effective. > bill davidsen (wedu@ge-crd.arpa) > {uunet | philabs | seismo}!steinmetz!crdos1!davidsen >"Stupidity, like virtue, is its own reward" -me Seriuosly, when people advocate doing things at run-time which we already know can be done effectively at compile time (once per program!) I wonder. The lessons from RISC and VLIW/TRACE seem to be: if you can move a runtime decision (like instruction scheduling) into the compiler not only does the hardware become simpler and hence faster, but that's one less thing to do at run-time. If we could build silicon to do a ``better and faster job ...'' then we pobably will be able to build one helluva compiler :-). David Callahan Rice Univeristy
mcg@omepd (Steven McGeady) (04/02/88)
Three weeks ago I posted an article with the above title, asking, among other things, that we discuss in this group a broader range of issues regarding processor architectures and their success than just (e.g.) conditions codes or not, scoreboarding or not, or UNIX-applicability or not. Other than a few small factual responses to the yield and silicon foundary question, none of the topics I suggested have been raised, and this discussion has devolved into another tedious RISC. vs. CISC debate! The questions I initially asked where (briefly): 1) Who manufactures MIPSco's silicon?, with details. This was partially answered, both by MIPSco and by others, with names of the foundries ("strategic partners") that MIPSco uses, and other information such as yield was understandably not present. 2) How does MIPSco keep apace of silicon technology? The argument was presented by MIPSco that their foundaries and strategic partners have competitive silicon technology which they can exploit. This begs the question to some extent of whether a foundary can be competitive with captive technology development, but that's likely to be a matter of opinion. 3) What is the availability of compilers, assemblers, debuggers) on VAXen, SUN's, and PC's? Absolutely NO ONE has addressed this. John Mashey politely did address it in private mail, but nobody else seems to care. John pointed out that MIPSco's compilers (a major part of their competitive posture) are available only on MIPSco boxes. This harkens back to the days of captive development systems for microprocessors, which I thought had been finally vanquished. I wonder why there is no discussion here of this. 4) What is the complexity of integrating a MIPSCo chip set into a system? What amount and kind of support HW is needed? I got a brief response in private mail from MIPSco saying "No harder than anybody else". Being a software type, I don't know all the right probing questions to ask, but surely there are others out there who either: a) have experience with this; or b) are curious and know the right questions. (E.g. memory bus i'face: how wide, deep, long, strong?; cache: how fast, expensive, hard, necessary?) 5) What does it take to make an architecture successful? I posited a (somewhat circular) answer to this that I thought would engender some discussion. I am disappointed. If you all ignore this message too, I will assume that my interest in these questions is perverse and crawl back into my hole. Finally, One respondent picked up on my reference to some "somewhat inelegant" aspects of the 386 (along with some inelegant aspects of the MIPSco chip). I pointed this out only to be (or appear to be) somewhat objective. Let there be no mistake: the 80386 pays my (meager) paycheck every month, and even though I'm working on unrelated architectures directed at different problem sets, and regardless of how I truely feel about the 386 architecture, which you won't be able to find out without knowing me much better, I am daily grateful to the 386 and its progenitors for the continued rise in the value of my Intel stock holdings. And, in case you hadn't figured it out, this and all other communications by me are my opinion alone, and in no way represent the opinion of anyone else, least of all my employer, and are presented without the explicit review or approval of that employer, something that will probably get me in big trouble someday. S. McGeady Intel Corp. mcg@iwarp.intel.com, mcg@omepd.intel.com, ...!intelca!omepd!mcg, ...!tektronix!ogcvax!omepd!mcg