davidson@SDSU.EDU (Craig Davidson) (07/15/89)
Just in case no one has noticed, Intel recently introduced a new processor
called the i-860. Since then, many companies have announced that they will
bring i-860 based systems to market. Some traditional 'Transputer companies'
are included in the list and so tales of mass abandonment of the Transputer has
spread through the rumour mills. The main reason given for the leaving of the
Transputer is usually reported as performance of the processor. It is time to
wake up and realize that it is not the processor performance that matters, but
the company selling it. The purpose of this diatribe is to try to shake loose
some people both inside Inmos and outside to help pressure Inmos to actively
fight for their market.
Anyone who is only involved in research and thinks that the commercial
realities do not apply to them, consider this: If the Transputer does not sell
well, it will be construed as a failure of the technology, not the marketing.
You will find it very difficult to get Industrial/Government research funding
for something which is a proven failure in the market.
Before going into this in much more depth I will let you know my bias. I have
been working with, or trying to work with, the Transputer and occam since 1984.
In my opinion the Transputer is elegant and probably the best design for large
parallel processing systems. Many companies have proven this, not only though
Transputer based machines, but the Ncube, Intel iPSC and Amatek hypercubes.
The occam language is elegant and powerful. I currently teach a Macintosh
version of Peter Welch's Occam and Transputer course here in the United States.
The response to occam has been very favorable and the attendees are generally
impressed with power of the language and how it makes them think about their
problems. I am pro occam and Transputer in my bias.
MYTH: "Everybody is designing in the i-860 over the Transputer since it is a
better mousetrap and gives them performance for their money."
FACT: In a graphics application, which Intel claims the i-860 is designed for,
a single i-860 performed about the same as 4 T800-20 processors with 4 cycle
RAM. The retail price of the hardware is the same and the i-860 is NOT
expandable.
NOTE: Unlike early users of the T-800 who said with glee that it was all Inmos
claimed and more, people who have used the i-860, or built systems
incorporating the chip have found it is not 50 times faster, or 10 times
faster, depending on your measurement, but only about 4-6 times faster than a
Transputer.
CONCLUSION: People are designing in the i-860 for reasons other than pure
performance. What is it?
To find out we must look at the companies involved (Inmos and Intel) and try to
find their similarities and differences. Some of these differences are based
on the Thorne-EMI or publicly owned Inmos (old Inmos) some are observations
about the Thomson-owned Inmos (new Inmos).
1. They both claim to be semiconductor manufacturers specializing in advanced
CMOS technology.
2. They both believe there is a market in the high-end workstation or personal
computer market and that replicated micro-computer processors is a viable
direction for supercomputers to go.
3. They both use Regis-McKenna for their Public Relations. RM is one of the
top firms in the field, responsible for Intel, Apple and other successful
company's PR. Intel follows their advice, Inmos seems not to.
4. Product Roll-out. When Intel introduced the i-860 it very carefully
targeted its audience and specified its target markets. Yes, it did fumble a
little by first calling the i-860 a coprocessor, but this is mostly forgotten.
If anyone has seen the videotape of the roll-out of the i-860 or read the
marketing propaganda, you may have noticed that the wording describing the
processor was carefully chosen. The language was chosen to target the
Transputer in many markets. Inmos should be pleased that Intel thought
Transputer worthy enough for this attack. Inmos should be worried and fight
back. Intel seems to have contacted many of the Transputer developers to try
to get them to develop with the i-860 and they have learned from Inmos'
mistakes.
Intel and Inmos are sold through the same distributors so all it needed to do
was let the distributors' sales reps know about the i-860 and ask 'Are any of
your present customers potential developers' The sales reps would immediately
think of anyone who is developing for the Transputer. Nothing as subversive as
stealing a customer list is involved.
To have a number of third-party products working at the roll-out Intel provided
detailed specifications of the product to developers ahead of time. I have
heard from reputable sources that Motorola distributed documents describing the
68040 architecture, instruction set, and preliminary interface timing a year
ago. Six months ago they supposedly had preliminary pinout documents
available. The 68040 is not even sampling until early next year, and then only
to select developers.
The T801-30, and T805-30 are supposed to be in fabrication now, according to
Inmos, and yet all that I have heard about the timing is that some of the setup
times are the same while other signals are 50% shorter than the T800-20. At a
proper rollout of the 30Mhz parts Inmos should be able to point to a list of
third-party hardware and software developers and say 'You can buy it from them
now.'
The design of the T810 was frozen before the OUG meeting last March and is
promised for 1990. The chip will have very different timing, different
instructions, and require different code optimizing schemes from the current
generation of Transputers. Inmos should be planning a rollout in 1990 as a
media event presenting a consumer product chip. What I mean by this is the
media should be impressed by the number of third party hardware products and
sophisticated software development environments as well as applications
available at product introduction. The public should get from the introduction
the impression that not only is a single T810 similar in performance to the
i860, but that from day one there are numerous products supporting it, a dozen
hardware developers SHIPPING boards, 10 C compilers, 5 or 6 development
platforms and 50 application packages. To do this Inmos has to start working
with the developers today. If not the chip will be passed off as a technical
curiosity and not really a serious contender.
5. Software support. In the beginning there was Inmos and occam. When enough
noise was made, C was developed and Inmos licensed it, selling it as 'Inmos C'.
Since then about a dozen different C compilers have been developed. Inmos now
develops a new C, rumored to be their own to compete with all the other C
compilers. Five different operating systems or environments of note have
developed, Helios, Express, Trollius, Idris and Linda, so Inmos is now talking
about yet another.
Intel, on the other hand, contracted with two well-known compiler companies,
Greenhills and I think Whitesmiths, to develop compilers as the chip was
developed. On announcement of the part, beta versions of the compilers were
shipping, one for UNIX and one for DOS. A similar tact was taken for the
Fortran compiler. If you want a C compiler, Intel will sell you the selected
one for your operating system environment at list price, or you can talk to the
developer - what Intel actually prefers. If you want the tools on another
platform, Intel will port it for you, for less than what you would have to pay
Inmos for the privilege of going through the pain yourself.
I personally do not like C. I also have heard from customers that they would
write in occam IF it would run on other processor and operating system
combinations. Portability and longevity of the code is far more important than
elegance, that is why Fortran is still around. (Personally I prefer Fortran to
C for many applications, if only the compilers were better...)
6. Development System. Intel requires you to use their 386 based UNIX
workstation for development until the compilers have stabilized. This is just
to make debugging of the tools easier since UNIX is not portable. After that,
Intel will not be in the development system/emulator market at all. They are
working with some third-party manufacturers to provide development systems and
emulators for the i-860. In the past Intel did make development systems, but
now they only make the system for the i-860 and can be believed when they say
it is only temporary.
Inmos claims it is only a chip manufacturer and yet prices its development
systems to compete with its developers. Inmos recently introduced lower
pricing and new Transputer modules to compete directly with third party
companies. In addition they have sold to Universities at 50% off their list
price knowing full well they were bidding against both their own distributors
and third party developers. This is worse in the new Inmos.
7. Evaluation units. Intel knows that to get parts designed in it must provide
a few free samples. NCR knows this, TRW knows this, Motorola, Hitachi,
Toshiba, AMD, National Semiconductor all know this, Inmos does not. In the
past Inmos could claim poverty, or an owner that did not understand
semiconductors, although I would have thought that a record company would
understand free samples. They are now owned by a large semiconductor
manufacturer and have bragged that they have a 4 million dollar marketing
budget and cannot think how they will spend it. It is still impossible to get
samples.
8. Comarketing. Everybody does it. This can range from showing third party
products in your advertisements and having third party companies in your booth
at trade shows to cooperative advertising money based on sales. After two
years of being told this was a good idea by Regis-McKenna, Inmos finally
invited companies into their booth at Siggraph last year. Usually the space is
free, since having developers showing products reflects on the quality of your
product and is a symbiotic advertising event. The third party provides
product, bodies, press releases and draws people to the booth, the host
provides booth space and coordinates publicity. All companies do this whether
they are a multi-billion dollar Apple or Intel or merely a small startup. It
is expected. Inmos charged $5000 for the service, four times the rate for the
floor space, and did not list the companies' products in the show catalogue.
To be fair, they did not list their own, either.
Intel provides all the traditional comarketing arangements. They have in the
past, they know it works, and they will in the future.
If we summarize the differences in what the companies do, not their products we
have:
INTEL INMOS
Actively pursue developers YES NO
Co-develop 3rd party software YES NO
Evaluation units YES NO
Compete with 3rd party
development software NO YES
Compete with 3rd part
development hardware NO YES
Provide tools to port to
other platforms YES NO
Co marketing YES NO
If this is what you saw, who would you develop for?
Inmos may claim that this is all changing under Thomson, but it has not yet
changed visibly for the better, only for the worse. In the short term Inmos
may realize better profits by competing with its customers, but in the end it
will chase them all off to use another chip and be left with no developers and
that will hurt everybody wanting to use Transputers and occam.
Intel realizes it needs developers and is actively working with them. IBM's
PS-2 floundered without third party developers, even though it has an annual
R&D budget which is about twice as large as the lifetime sales of Inmos, or
twice the current sales of Thomson, and can theoretically develop anything that
is needed. If even IBM cannot sell a product without third party developers,
why should Inmos think it can?
Craig Davidsonneighorn@qiclab.UUCP (Steven C. Neighorn) (07/17/89)
In article <8907150609.AA01452@sdsu.edu> davidson@SDSU.EDU (Craig Davidson) writes: <text deleted> >Before going into this in much more depth I will let you know my bias. I have >been working with, or trying to work with, the Transputer and occam since 1984. >In my opinion the Transputer is elegant and probably the best design for large >parallel processing systems. Many companies have proven this, not only though >Transputer based machines, but the Ncube, Intel iPSC and Amatek hypercubes. What do you mean by this last sentence? How have Ncube (which uses a custom- made proprietary chip), Intel (which used the 80286 and now uses the 80386 chip) and Amatek proven anything? Are you saying these companies in not using the Transputer are proving the Transputer is "elegant and probably the best design for large parallel processing systems" ? >The occam language is elegant and powerful. I currently teach a Macintosh >version of Peter Welch's Occam and Transputer course here in the United States. >The response to occam has been very favorable and the attendees are generally >impressed with power of the language and how it makes them think about their >problems. I am pro occam and Transputer in my bias. Perhaps, but I can think of two cases that might lead to different opinions about Occam. First of all, Cogent Research, which makes a Transputer-based workstation called the X/TM, uses Linda (C-Linda, kernel Linda, etc) as its primary programming language. If the Transputer and Occam are such a wonderful fit, why has Cogent Research jumped on the Linda bandwagon? Secondly, Floating Point Systems didn't have what most would call successful market penetration with their T Series machines, also based on the Transputer, but offering Occam as its first programming language. I have talked to many former Floating Point employees (and believe me, around here there are a ton of 'em) and most told me they felt forcing Occam down customer's throats (let alone their own) was a mistake. They didn't like the language. In the August 1986 issue of the "Communications of the ACM" in an article by Karen A Frenkel, the FPS T series machine is discussed. The article talks about Kenneth G. Wilson of Cornell University, who helped with the design, not being happy with Occam and expecting Unix to be running on the machine in a few months. Charles L. Seitz of Cal Tech didn't seem to fond of Occam either, declaring it "... the Fortran of parallel processing..." I certainly don't want to start another processor war, but I am curious as to what was meant by Transputer being the best design for parallel machines. -- Steven C. Neighorn !tektronix!{psu-cs,nosun,ogccse}!qiclab!neighorn Sun Microsystems, Inc. "Where we DESIGN the Star Fighters that defend the 9900 SW Greenburg Road #240 frontier against Xur and the Ko-dan Armada" Portland, Oregon 97223 work: (503) 684-9001 / home: (503) 641-3469
wm@CSE.OGC.EDU (Wm Leler) (07/18/89)
> Perhaps, but I can think of two cases that might lead to different > opinions about Occam. First of all, Cogent Research, which makes a > Transputer-based workstation called the X/TM, uses Linda (C-Linda, > kernel Linda, etc) as its primary programming language. If the > Transputer and Occam are such a wonderful fit, why has Cogent Research > jumped on the Linda bandwagon? I don't think our choice of Linda had much to do with whether occam is elegant, precise or whatever. It is a mistake to assert that there is one correct or best language for parallel programming (or sequential programming for that matter). I know many satisfied occam programmers (besides Craig!), and can think of applications where occam would be a better language than C and Linda. Writing an operating system is not one of them. And that ignores the commercial realities of occam in North America. I would also like to point out that Linda is not a language, it is (mainly) a communication paradigm. There is no reason why one couldn't build occam-Linda. > I certainly don't want to start another processor war, but I am curious > as to what was meant by Transputer being the best design for parallel > machines. The transputer was the first microprocessor designed for parallel processing. It is simple, elegant, and powerful. Steve's comments are well taken, but I would like to see more discussion about what (I believe) Craig's posting was all about -- that with a little hype and hoopla, Intel is taking the market away from Inmos, and Inmos seems to be doing everything in its power to help them. Wm Leler Cogent Research p.s. I've heard a rumor that Inmos/Thompson has recently dropped Regis-McKenna, and will be doing their own PR. p.p.s. To their credit, I believe Inmos has offered free booth space at SIGGRAPH this year, so maybe they are starting to wake up?
les@unicads.UUCP (Les Milash) (07/18/89)
In article <2238@qiclab.UUCP> neighorn@qiclab.UUCP (Steven C. Neighorn) writes: >I certainly don't want to start another processor war, but I am curious >as to what was meant by Transputer being the best design for parallel >machines. I will be happy to explain why people who make this claim do so, although I won't "argue" with you in the usenet "style" about it. SCN and anybody else are welcome to e-mail me if you'd like to start round 1. Les Milash RISC (Reduced Investor Satisfaction Corp :-)
davidson@SDSU.EDU (Craig Davidson) (07/18/89)
>What do you mean by this last sentence? How have Ncube (which uses a custom- >made proprietary chip), Intel (which used the 80286 and now uses the 80386 >chip) and Amatek proven anything? Are you saying these companies in not using >the Transputer are proving the Transputer is "elegant and probably the best >design for large parallel processing systems" ? What I meant was the CSP (Communicating Sequential Processes) model of programming distributed memory systems. This is directly implemented in occam and in hardware by both the Ncube and Transputer processors. All the hypercube machines and their operating systems use this parallel processing model as well. Most versions of C, Fortran and Pascal that I am familiar with on distributed memory machines also have followed the CSP model. Even Ada (through rendezvous etc) and Linda are CSP based systems. It is the CSP architecture which is elegant. >Secondly, Floating Point Systems didn't have what most would call >successful market penetration with their T Series machines, also based on >the Transputer, but offering Occam as its first programming language. I have >talked to many former Floating Point employees (and believe me, around here >there are a ton of 'em) and most told me they felt forcing Occam down >customer's throats (let alone their own) was a mistake. They didn't like the >language. In the August 1986 issue of the "Communications of the ACM" in >an article by Karen A Frenkel, the FPS T series machine is discussed. The >article talks about Kenneth G. Wilson of Cornell University, who helped >with the design, not being happy with Occam and expecting Unix to be >running on the machine in a few months. Charles L. Seitz of Cal Tech didn't >seem to fond of Occam either, declaring it "... the Fortran of parallel >processing..." Floating Point Systems T-series market penetration would have been as bad if they had offered C without reals or multi-dimensional arrays, with no assembler or instruction set reference manual, no standard numeric libraries and no operating system. The example of Floating Point Systems is also a reaffirmation of what I was trying to say, that is Inmos has been reluctant to work with third party developers to produce industry standard languages such as C or environments close enough to UNIX to be used. To their credit they are much better now. I was also trying to point out that Intel is learning faster than Inmos. Charles L. Seitz's comment that occam was "... the Fortran of parallel processing" is not necessarily a negative comment. Occam is a start, it has the control structures necessary for parallel processing and it is something to build on. Even Tony Hoare and David May would agree with Charles Seitz and do not claim that occam is the ultimate parallel language, only a useful beginning. As to Fortran, last I looked vendors are now starting to offer Fortran-88, how many languages have, or will, survive for over 30 years with the popularity of Fortran. The Ncube, Intel, and Amatek hypercubes all offer Fortran, why not the "Fortran of parallel processing?" Since Charles Seitz was brought up, does anyone have the article reference where he describes the hypercube chip he is designing? re: Wm Leler's p.p.s. about booth space at Siggraph. Last I heard Inmos had planned to fill its booth with its own products. They had offered free space in smaller shows earlier this year. Craig Davidson
siemon@exunido.uucp (Peter Siemon) (07/21/89)
What Craig Davidson is perfectly o.k. But his conclusion is wrong. Obviously INTEL is the better company, has the better product, the better marketing, the better service and the better support. I cannot argue about the bad things he said about INMOS as we have a different distributor here. But things seem to be very similar at this end of the world: prices and support, availability and documentation - mail me for my personal remarks (or better not). So - if INMOS looses. I do not care. The idea of parallel processing has not been invented by them. And it might be perfected by someone else!
chapman@compsci.bristol.ac.uk (Paul Chapman) (07/22/89)
In a recent message Peter Siemon writes :- >> So - if INMOS looses. I do not care. The idea of parallel processing has >> not been invented by them. And it might be perfected by someone else! If you feel so negative about Transputers why are you using them? Or are you just waiting for something better to come along? BUT if you are doing something worthwhile shouldn't you be concerned that your supplier may be in danger of beeing squeezed out of business and everything you have done becoming obsolete. No doubt anything you have done can be transfered to another device, but I for one have invested a great deal of time in Occam and have no desire to rewrite it all. >> What Craig Davidson (says) is perfectly o.k. But his conclusion is wrong. >> Obviously INTEL is the better company, has the better product, the better >> marketing, the better service and the better support. NO obviously about it. I think Craig's remarks are very much to the point. The British have a long tradition of developing good ideas and not being able to capitalise on them and I see no reason for the Transputer to be any different. The brutal truth is that Our Beloved (?!) Prime Minister's attempts to Americanise our economy by placing it at the mercy of market forces has not worked because there is not, and has not been for far too long, the degree of investment (both in terms of capital and management expertise) that is required in our industry -- the abysmal under-funding of Inmos by Thorn-EMI is just one more testimony to that. I have heard more than one person express the view that Inmos survives despite, rather than because, of the way it is run. On top of all this, US firms have the advantage of a huge home market on which to develop their products. The question is whether 1992 will come in time to save the Transputer!! Unfortunately there is not much we can do on this net to change Inmos company policy. However, it would be interesting to hear from those in The Company who get this stuff, to see what they think -- or is this to contentious to handle without fear of unemployment !?! Paul Chapman. (Dept. Computer Science, Bristol University, U.K.)
neighorn@qiclab.UUCP (Steven C. Neighorn) (07/24/89)
In article <8907181536.AA13030@sdsu.edu> davidson@SDSU.EDU (Craig Davidson) writes: >What I meant was the CSP (Communicating Sequential Processes) model of >programming distributed memory systems. This is directly implemented in occam >and in hardware by both the Ncube and Transputer processors. All the >hypercube machines and their operating systems use this parallel >processing model as well. Most versions of C, Fortran and Pascal that I am >familiar with on distributed memory machines also have followed the CSP model. >Even Ada (through rendezvous etc) and Linda are CSP based systems. It is the >CSP architecture which is elegant. Then I wish you would have said CSP was the concept that was elegant and probably the best for large parallel processing systems. But I'm not sure I entirely agree with your conclusions about CSP being the basis for machines other than Transputer machines, though it may be a problem with semantics more than anything else. I don't look at hypercube's loosly coupled MIMD architecture or Ncube's architecture as being specifically CSP-based, as defined by Hoare. Yes, processes run in distributed memory partitions, and yes, those processes communicate with each other at varying speeds, depending on implementation, but do they fully fit the CSP model? I look at CSP less as a generalized description of parallel computation and more as a mathematically sound paradigm that covers conventional sequential programming within a communicative parallel environment. The beauty of CSP is the mathematics, the handling of nondeterminism, and the application towards systems of processes. I fail to see how the key points and beauty of CSP are fully applied in the hardware or software you discuss above, except for the Transputer and Occam. Occam follows the CSP guidelines closely, and the Transputer is a natural home Occam. But the others? If I'm way off base here, I'd like to find out in the friendly, learning environment of USENET rather than the hostile accommodations of a post-final exam computer science debate at the local watering hole. I also don't understand what you mean when you say that most versions of C, FORTRAN, and Pascal on distributed memory machines follow the CSP model. How so? Ada is obvious to me, with rendezvous as you mentioned, along with priority. >Floating Point Systems T-series market penetration would have been as bad >if they had offered C without reals or multi-dimensional arrays, with >no assembler or instruction set reference manual, no standard numeric >libraries and no operating system. The example of Floating Point Systems >is also a reaffirmation of what I was trying to say, that is Inmos has been >reluctant to work with third party developers to produce industry standard >languages such as C or environments close enough to UNIX to be used. To >their credit they are much better now. I was also trying to point out >that Intel is learning faster than Inmos. I don't like to see any chip and/or its manufacturer die because its marketing department doesn't know what the Hell its doing. Chips should die because they are outclassed, used up, or beaten by something better on the market. If Intel's marketing is better than Inmos', then I hope Inmos gets their marketing people gets on the stick. If Intel's chip is better than Inmos', then I hope Inmos gets their engineers on the stick. >Charles L. Seitz's comment that occam was "... the Fortran of parallel >processing" is not necessarily a negative comment. Occam is a start, it has >the control structures necessary for parallel processing and it is something >to build on. Even Tony Hoare and David May would agree with Charles Seitz >and do not claim that occam is the ultimate parallel language, only a >useful beginning. As to Fortran, last I looked vendors are now starting >to offer Fortran-88, how many languages have, or will, survive for over >30 years with the popularity of Fortran. The Ncube, Intel, and Amatek >hypercubes all offer Fortran, why not the "Fortran of parallel processing?" >Since Charles Seitz was brought up, does anyone have the article >reference where he describes the hypercube chip he is designing? Lest I also dip into the realm of good/bad FORTRAN arguments, let me just say that FORTRAN-88 and 30 years of popularity are not necessarily a good thing. In fact, some people who are trying to move programming languages into the 90's might even say they were a bad thing. FORTRAN is popular because it *was* popular. No one wants to go back and tear apart that hundred-thousand line simulation and convert it into something a tad bit more modern. It has a long history, and it was right there at the beginning, and '88 makes some nice improvements, but FORTRAN and goodness just on the basis of longevity aren't going to wash with me. Why do companies line Ncube, Intel, and Amatek offer FORTRAN? Because they would be foolish not to. I mean someone's got to handle those old versions of QUANC8. I know there is a place for the language, but tell me are there any inherent properties in FORTRAN that make it suited for parallel programming *besides* its history? I can't say whether Charles Seitz's comment really was meant as an insult or a compliment. In the context of the article, it was definitely an insult. -- Steven C. Neighorn !tektronix!{psu-cs,nosun,ogccse}!qiclab!neighorn Sun Microsystems, Inc. "Where we DESIGN the Star Fighters that defend the 9900 SW Greenburg Road #240 frontier against Xur and the Ko-dan Armada" Portland, Oregon 97223 work: (503) 684-9001 / home: (503) 641-3469