Meiko Scientific

» see meiko.com as it lived, on the wayback machine

~ the obit ~

In 1985 the management at Inmos wanted to hold the transputer back a while longer. Six of the people working on it disagreed strongly enough to quit: Miles Chesney, David Alden, Eric Barton, Roy Bottomley, James Cownie and Gerry Talbot. They stayed in Bristol, called the new company Meiko after a Japanese word for well engineered, and by July they had a rack of experimental 16 bit transputers doing something visible on the floor at SIGGRAPH in San Francisco. The company's own account puts nine weeks between the resignations and the demonstration.

The product that came out of it in 1986 was the Computing Surface, built on the 32 bit T414. The name was literal. You bought a crate of transputers and wired them into whatever shape your problem wanted, because each chip carried its own memory and four communication links, and was designed from the start on the assumption that you would never buy just one. Universities took to it hard, and the biggest of the early installations was the Edinburgh Concurrent Supercomputer. By 1990 Meiko had shipped over three hundred systems and employed 125 people, which for a British computer company founded out of a walkout was a genuine business rather than a research project.

The CS-2 arrived in 1993 and moved to SuperSPARC and hyperSPARC processors with vector units available as an option. The largest one ever built had 224 processors and went to Lawrence Livermore. It was a good machine landing in a bad year. Racks of ordinary workstations lashed together with ordinary networking were becoming the cheap answer to the same problems, and the specialist parallel vendors on both sides of the Atlantic were running out of customers at the same time. Meiko ran into financial trouble in the middle of the decade and stopped trading in 1996.

There is one route by which Meiko silicon could plausibly have reached me, and it has the company's name nowhere on it. Meiko had hired Fred Homewood and Moray McLaren, both of whom had worked on the transputer's floating point unit, and the two of them designed a better one. It was aimed at Intel's 80387 instruction set until the litigation around that part made the plan commercially hopeless, and a conversation with Andy Bechtolsheim during a visit to Sun turned it toward SPARC instead. LSI Logic fabricated it for the SPARCstation 1. Sun eventually licensed the design outright and used it across several generations of MicroSPARC parts, for a one off payment and a full Solaris source license. Any Sun workstation sitting in a Turkish university lab in my student years had a Bristol floating point core in it, sold once, badged by somebody else forever.

~ the rap sheet ~

~ the chip that never came ~

The CS-2 was supposed to be tied together by the T9000, the transputer generation Inmos had been promising since 1990. In 1990 Meiko started an internal project codenamed Elan-Elite as a hedge against that not arriving. The T9000 then slid so far and so publicly that the hedge became the only option on the table, which is a strange position for a small company to find itself in: the thing you built in case your supplier failed is now your entire product.

Elan was the adapter, a SPARC based communications coprocessor sitting on the Sun MBus with two bidirectional links at 50 MB per second. Elite was an eight way crossbar switch you built the packet network out of, with some adaptation for load and priority. Both were gate arrays and GEC Plessey fabricated them at Roborough near Plymouth in 1993. For a company of 125 people to design its own network silicon because its chip vendor could not deliver is the kind of thing that reads as heroic in a history and as a cash flow problem at the time.

~ what quadrics did with it ~

In mid 1996 the technical staff and the technology went to Quadrics Supercomputers World, a joint venture put together by Alenia Spazio of Italy, with offices in Bristol and Rome. Elan and Elite became QsNet, and QsNet spent the following decade as one of the two or three interconnects that serious cluster builders argued about, holding together some of the largest machines at the American national laboratories. The computer died in 1996. The wiring survived it by more than ten years and went into machines far larger than anything Meiko ever shipped.

The people scattered usefully. Eric Barton, a founder and Meiko's director of software development for a decade, spent the 2000s on Lustre, the parallel file system that most of the world's large clusters store their data on, running the engineering for it at Cluster File Systems and then at Sun after the 2007 acquisition, then co-founding Whamcloud in 2010 to keep it open source when Oracle lost interest. The Bristol walkout of 1985 bought eleven years of supercomputer company, and left behind an interconnect and a file system that are both still load bearing.

@}-->--

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

« previous  ·  random  ·  next »