🚀 go-pugleaf

In article <9002201905.AA24148@polaris.cs.uiuc.edu> north@POLARIS.CS.UIUC.EDU (Michael J. North) writes:
>
>imho, there are lots of reasons for this,
>
>	- the transputer came out 5 years too late for it's market

Well, not five years, but probably a couple. By the time many people had
even heard of them, there were plenty of other processors which could
match the speed, which was one of the main advantages. I think (hope ?)
that it may eventually turn out that they were in fact too EARLY for the
market.

>	- the transputer is designed as a CSP engine, C and other languages
>	    were delivered as an afterthought to Occam,  i also
>	    seem to remember the compiler writer's guide wasn't
>	    available to mere mortals

Well, I had one on my desk for a while. I don't think they're that difficult
to get hold of. On the other hand, as you say, the internal design leaves
a little to be desired from many points of view. I would imagine that 
originally OCCAM was the only language you were expected to use.

>	- no mmu.  virtual memory is a wonderful thing, try it some
>	    time.

Personally, I think this is the biggest problem from the workstation point
of view. I have tried to design an MMU which could be attached to the 
beast, but the amount of hacking which would be required simply isn't
worth it. Hopefully the H-1 will fix this, 'cause I just don't see how
you can run a useful operating system without virtual memory, unless you
can afford 64 megabytes on each processor.

>	- i suppose there is some north american snobbery about the 
>	    chip being british.  just look how much the american
>	    market uses japanese microprocessors (the japanese 8086
>	    look alike comes to mind but there were a ton of lawsuits
>	    i seem to remember)
>	- inmos didn't have the world's greatest balance sheet,
>	    concern over whether your supplier would be around
>	    next month would bother any rational company who wanted
>	    to generate a profit (you did mention something
>	    about wild-eyed entrepeneurs).  americans are notorius for
>	    the quick buck.......

Yes, many companies are reticent about using any chip which is only
manufactured by one supplier. Unfortunately I seriously doubt that Inmos
would consider allowing ANY other company to build transputers, even 
under licence.

>	- this was rehashed in this list a while back when the
>	    topic was the i860, but anyone who introduces a transputer
>	    based product may find themselves competing with inmos.
>	    inmos also supports an operating system (i don't remember which
>	    one), some compilers, and i think even builds a vme board.
>	    you end up competing with your supplier -- this has
>	    the makings of a no win situation.

Yeah. Inmos seem mainly to want to sell their own boards, compilers,
etc, etc rather than selling the chips to other companies. I used to 
work for an avionics company who wanted to put them into jet fighters, and
the hassle of actually getting raw chips was, well, somewhat excessive.

On the other hand, a friend of mine is involved in a contract with Inmos
writing ADA compilers and apparently has more transputers in, on or under
his desk than he knows what to do with (engineering samples). 

>	- another topic is that of parallelism itself, why should you take
>	    the effort of getting your application(s) to run in parallel
>	    when the next generation uni-processor might be able to
>	    run it much faster anyway.  parallelism is often "non-trivial".
>	    uni-processors may not be able to do this for too much
>	    longer, but for the immediate future it's likely that they will.
>	    

Well, our embedded systems used 12 (mixed T4/T8) processors, that's about 
120 MIPS and about 6 MFlops. It's still going to be a couple of years before a
single processor can match that, even allowing for the ratio between 
transputer MIPS and CISC MIPS. In my experience, parallelism often seems to
be much easier to implement after a little thought than it may first appear,
although it may turn out that the parralel algorithm is actually SLOWER
than serial, due to the slow serial links on the current processors
(only twice as fast as EtherNet after all (1/2 B-)). Certainly the 
systems I was working on were being designed using multi-port RAM,  
rather than using links. 

That said, there are many applications where parallelism is just a non-starter.

>
>the transputer is really an elegant chip.  it just had a bum rap.
>
>
>
>michael j north
>north@polaris.cs.uiuc.edu
>tapestry laboratory for parallel systems

From what I've been told, it wasn't really designed to be a general purpose
processor, but was intended largely for embedded systems (someone out there
must be able to comment on that one). If the H-1 specs I've seen are
correct (100 MIPS, 20 MFLOPs, 20MByte/sec links) then hopefully it will
take off this time round.
 
Whatever the case, the transputer deserves a lot more attention. It could
be that it's market is only just appearing now as networks become more 
commonplace.

Finally, to add my suggestion :

	- Three different instruction sets on processors of the same
family !

Obviously, the different processors are designed for different applications,
but even so, it adds a lot of extra hassle to development in a mixed
system.