Trilogy Systems

~ the goodbye ~

Trilogy stopped existing in October 1985, when I was one year old. Everything below comes off the record rather than out of anything I could remember, and the record here is unusually complete, because a company that raises $230 million and ships nothing gets written about at length.

Gene Amdahl was the reason the money moved. He had been the chief architect of the IBM System/360, left IBM in 1970, and built Amdahl Corporation into a business that sold mainframes plug-compatible with IBM's own and undercut them. That is close to the strongest track record anybody in computing had in 1979. He left Amdahl Corporation that August, and by 1980 he and his son Carl had incorporated a new company under the name ACSYS, soon renamed Trilogy Systems.

The pitch was wafer-scale integration. Rather than fabricate hundreds of small dies and wire them together across boards, print the whole logic assembly onto one enormous die and let the connections be part of the same photolithography. Trilogy's target was a chip 2.5 inches on a side at a moment when a normal one was a quarter of an inch, holding 20,000 to 40,000 gates in high-speed ECL, replacing roughly 120 conventional integrated circuits. The machine built out of them was supposed to run two to four times faster than the quickest thing IBM sold, at a fraction of the assembly cost.

Investors put in $55 million through a limited partnership, then $30 million of venture money, then $60 million more in a 1983 public offering. The total came to somewhere north of $230 million, which in 1983 dollars made Trilogy one of the largest startup bets ever placed on an unproven manufacturing process.

~ the chronicle ~

~ the chip ~

Defects on a silicon wafer land more or less at random, so the odds of a die coming out clean fall off sharply as the die gets bigger. At 2.5 inches square, the probability of a perfect one is effectively zero. Trilogy knew this and designed around it: build redundant copies of every circuit block, expect 300 to 400 defects per chip, and route around the dead ones after test.

The engineering was sound in outline and brutal in detail. Metal deposition kept pulling the wiring loose until a nickel layer fixed it. Chemicals leaked through the process. Each turn of the design took about eight months to come back from the fab, which meant roughly one learning cycle a year on a problem with dozens of unknowns. In December 1983 a prototype test ended with two microscopic wires touching, the whole circuit shorting, and the block of silicon heating until it glowed red.

None of it was wrong, only forty years early. Wafer-scale parts with redundancy and post-test routing are how Cerebras builds AI accelerators now, on tooling Trilogy could not have imagined.

~ the money ~

The company also absorbed the kind of damage no schedule accounts for. Clifford Madden, the third founder, died of a brain tumor, and Gene Amdahl was hurt badly enough in a car accident to be out of the building. A winter storm got into the fabrication plant while it was still under construction.

In June 1984 Trilogy announced it was abandoning the IBM-compatible mainframe entirely and would sell its technology to other computer makers instead. Carl Amdahl left that same month and told the New York Times, in effect, that the thing had been harder than anyone had allowed for. Gene Amdahl stepped back from the chief executive role by mid year and Henry Montgomery took it. In June 1985 the board announced it would spend what was left on a majority stake in Elxsi, a San Jose minicomputer startup, and the deal closed that October for roughly $52 million in Trilogy stock. By the end of 1985 all remaining chip development was stopped.

~ what was left ~

The merged company kept Elxsi's name and Trilogy's cash. It stayed in computers until 1989, when it walked out of the market for the same reason everyone else did, which was that the personal computer had made the departmental minicomputer pointless. Gene Amdahl left the same year and founded Andor Systems, which went bankrupt in 1995.

The corporate shell outlived all of it. Elxsi sold the maintenance business in 1989 and in 1991 went into two lines of work with no relationship to anything above: restaurants, and video equipment for inspecting sewer pipes. It is still in those businesses. The name went the other way. Tata had been an Elxsi investor and kept the word for Tata Elxsi, incorporated in 1989 and still trading on the Indian exchanges.

@}-->--

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

« previous  ·  random  ·  next »