Inmos

~ the eulogy ~

Inmos was founded in July 1978 by Iann Barron with two American semiconductor people, Richard Petritz and Paul Schroeder, and it was funded by the British state through the National Enterprise Board with an opening £50 million. The premise was industrial policy rather than a business plan. Britain had no volume chip manufacturer and the government decided to buy one into existence, with a design operation in Bristol and a fab in Newport, South Wales.

The product it is remembered for is the transputer, designed by a team under David May and Robert Milne, in production from 1985 with the T212 and T414. A transputer was a processor with memory and four communication links on the same piece of silicon, built on the assumption that you would not buy one. You would buy a great many and connect them into a mesh, and the parallelism would come from the wiring rather than from anything clever inside a single chip.

Ownership changed twice and neither change was about the technology. Thorn EMI bid £124.1 million for the state's 76 percent holding in July 1984 and raised it to £192 million before the deal closed that September. SGS-Thomson, now STMicroelectronics, bought the company in April 1989. In December 1994 Inmos was fully absorbed and the name stopped being used, after roughly £211 million of public money over its lifetime.

I work on computer vision at the edge, which means getting parallel arithmetic to run on small chips with a fixed power budget. That is close to the thing Inmos was selling in 1985 and could not make anyone want. The difference is not the idea. It is that the tools got there, forty years later, and that the market eventually contained enough problems shaped like this one to be worth building for.

~ the dossier ~

~ what the money bought ~

The transputer came with its own language. Occam was built on the same model as the hardware, with processes that communicated over channels and no shared memory to argue over, derived from Tony Hoare's work on communicating sequential processes. Hardware and language were designed as one thing, so a channel in your program could be a link on the board or a link between chips and the code did not need to know which.

This was coherent and it asked a lot. Programmers had to think in concurrent processes at a time when almost every working engineer thought in sequential C, and the payoff only arrived if you restructured the problem to suit the machine. The transputer found real customers in supercomputing and embedded control, and in a few strange corners like the Atari ATW800 workstation, but it never crossed into general purpose computing where the volume was.

~ the postmortem ~

The T9000 was supposed to be the generation that made the argument unanswerable, and it did the opposite. The design was budgeted at 2.2 million transistors and grew as the engineers added the circuitry needed to track dependencies across its eight-instruction pipeline. It was promised for the first quarter of 1992 at 50 MHz and slipped to the fourth quarter of 1993 at 40 MHz, slower than the original target and two years late.

Two years was the whole margin. Conventional processors were compounding through the early 1990s at a rate that made a late parallel chip pointless, because by the time it arrived a single ordinary CPU could do the work of the mesh you were being asked to design around. The performance case died first and the price case went with it. SGS-Thomson had no reason to keep a brand attached to a strategy that had lost, and the name went in December 1994.

~ what got buried with it ~

The transputer was right about the destination and wrong about the schedule by about twenty five years. Every phone now runs heterogeneous cores passing work between each other, and the message passing model Inmos built into silicon is the standard way large machines are programmed. None of that traces back to Inmos as an ancestor. It arrived independently once the economics forced it.

David May went to the University of Bristol and the design culture stayed in the city, which is part of why Bristol still has a chip industry. The public money is harder to score. £211 million bought a fab and a genuinely original architecture, along with a generation of engineers who kept working after the company's name was switched off. It did not buy Britain the semiconductor industry the National Enterprise Board was shopping for in 1978.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »