ETA Systems

~ the last words ~

Control Data had invented the supercomputer business and by the early 1980s it was losing it to a company founded by one of its own former engineers. William Norris, who had run CDC since 1957, decided the problem was his own management layer. The fix was to put the supercomputer work somewhere it could not be reached. ETA Systems was established in August 1983 as a separate company with one instruction, which was to deliver a ten gigaflops machine by 1986. Neil Lincoln, chief architect of the Cyber 205, went across to design it.

The bet underneath the ETA10 was CMOS. Every fast machine in the world at that point was built from bipolar logic, which switched quickly and ran hot enough to need elaborate plumbing. ETA went to Honeywell instead and had a 20,000 gate CMOS array built, the ALSI-20K, which let an entire central processor sit on one board. Nobody else in the industry thought this was how you built a supercomputer. It took about a decade before everyone was doing it.

The first processor left St. Paul on 31 December 1986 and installation started at Florida State University on 5 January 1987. The machine was formally announced that April at the customer's site, which is the sort of thing a company does when it needs the announcement to be about a working installation. Sales were real but thin, and the losses were not. ETA burned roughly a hundred million dollars a year against a parent company that was already bleeding, and Norris himself retired in January 1986 with the machine still unfinished.

On 17 April 1989 Control Data closed ETA Systems and terminated its staff. Seven liquid cooled and twenty seven air cooled machines were installed at that point. Florida State learned about it the same day, in the middle of formal negotiations with CDC over a hardware upgrade for the machine it already owned.

~ the record ~

~ the cold box ~

CMOS at room temperature was slower than the bipolar parts it was replacing, so ETA made the room colder. On the E and G models the processor boards were submerged in liquid nitrogen at 77 kelvin, around minus 300 Fahrenheit, which more than doubled the speed of the same silicon. The G model reached a seven nanosecond cycle and a peak of 10,289 MFLOPS, and it is still the only production supercomputer ever built that had to be kept cryogenic to hit its numbers. The air cooled P and Q models ran the same logic at 24 and 19 nanoseconds.

~ the software never came ~

The hardware arrived years before anything to run on it. Through 1987 the machine at Florida State had no operating system at all and ran in monitor mode, with programs loaded one at a time from workstations wired to the side of it. EOS, the operating system ETA had written from scratch, showed up in early 1988 and could not keep more than two users on a processor without falling over. A Unix System V port released on 3 October 1988 was the first thing the hardware ever ran that behaved, roughly twenty one months after the first customer took delivery.

The cost of that gap shows up in the benchmarks. In October 1987 an ETA10-E set a world performance record, and its LINPACK result was 52 MFLOPS against a machine rated in the thousands. Customers were not measuring the silicon. They were measuring how much of it the software could reach, and for two years the answer was almost none of it. Sites with deadlines bought Cray machines that worked, and most of them did exactly that.

~ the argument outlived the company ~

The first hand accounts from inside ETA carry a detail that is hard to shake. On the day the doors closed, the sales organization had just reached agreement to deliver more than forty systems to a customer in Australia, and shortly afterward the US government told Control Data it was ready to buy, by which point there was no ETA Systems to buy from. The manufacturing line, including the plating process developed in house for the boards, was dismantled. CDC gave some of the surviving machines to high schools through a program called SuperQuest. Florida State traded its ETA10-G back for a Cray Y-MP and had the replacement running by April 1990.

What stays interesting is that ETA was right about the only thing that mattered technically. Every processor I have ever owned is CMOS, down to the boards on my desk this morning, because bipolar logic lost the argument completely within about ten years of ETA being closed for making it too early. The company that saw it first died anyway, and not for seeing it. It died because a machine with no operating system is a very expensive space heater, and because the parent that spun it out to escape interference kept the controlling stake and the habit of interfering.

@}-->--

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

« previous  ·  random  ·  next »