Amdahl Corporation

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

~ the obit ~

i met this name as a law before i ever met it as a company, which is the normal order for anyone who came to computers through a Turkish engineering faculty in the 2000s. Amdahl's Law is on the whiteboard in the second or third week of computer architecture: if a fraction of your program refuses to run in parallel, that fraction sets a ceiling on how much faster you can ever make the whole thing, no matter how many processors you throw at it. It is one of the few results in the field that stays true as hardware changes underneath it. Nobody in that room mentioned that the man had also run a company, and the company had been dead for years by then anyway.

Gene Amdahl was the chief architect of the IBM System/360, which is the machine that made IBM the thing it became. He left in 1970 and incorporated Amdahl Corporation in Sunnyvale on 19 October of that year, on a thesis that was slightly rude and entirely correct. IBM's customers were not locked in by IBM's hardware. They were locked in by the operating system and the application code they had written on top of it. Build a processor that runs that software untouched and you can sell against IBM on price and performance alone.

The first machine, the 470V/6, went to NASA's Goddard Space Flight Center in June 1975. It ran roughly twice as fast as IBM's contemporary 370/168 for about the same money and took up a third of the floor space. Amdahl's processors were air cooled, which sounds like a detail until you remember that IBM's large systems of that era wanted chilled water plumbed into the machine room. The term for what Amdahl sold was plug compatible, and for a while the company more or less owned the phrase.

At its best it held about 24 percent of the IBM-compatible mainframe segment and somewhere near 8 percent of the worldwide mainframe business, which for a company competing directly with IBM's core product is a remarkable place to have ended up. Fujitsu had put money in at the very beginning, five million dollars when there was nothing to see, and it kept buying. Gene Amdahl himself left in 1979 to build a wafer-scale processor at Trilogy Systems, which did not work.

The end was slow and then administrative. Mainframe revenue slid through the early 1990s as the work moved onto smaller machines, and Amdahl bought a Canadian consultancy and tried to become a services company instead. In July 1997 Fujitsu agreed to pay around $850 million for the 58 percent of the shares it did not hold. The tender offer and merger completed on 19 September 1997. Late in 2000 the company announced it would not build a 64-bit IBM-compatible system, which was the end of the hardware, and the name itself thinned out of Fujitsu's American operations over the next few years without any announcement at all.

~ the rap sheet ~

~ the machine that made ibm move ~

The competitive effect was larger than the market share suggests. Before Amdahl, a large IBM account had no credible alternative that did not involve rewriting everything it owned. After Amdahl, the same account could put a competing quote on the table and mean it. IBM's pricing moved and its upgrade cycles shortened, because there was finally somebody to lose a deal to.

This is also why the company could never win outright. Its product was defined by another company's architecture, so IBM set the roadmap and Amdahl followed it closely enough to stay compatible. Every generation was a race to match a specification somebody else had published, and a supplier who exists inside a rival's design decisions has a ceiling on what it can become. Amdahl found that ceiling around a quarter of the segment and stayed under it for the rest of its life.

~ the law and the company ~

The law came first. Gene Amdahl presented the argument at a conference in 1967, three years before the company existed, and it was aimed at people who were claiming that enough parallel processors would solve any performance problem. The arithmetic says otherwise. Adding hardware does not shrink the serial fraction of the work, so that fraction ends up being the whole cost. Past a point the curve flattens no matter how much you spend.

That result has outlived everything else with his name on it by a wide margin. A 470V/6 sits in a museum now and the corporation is a line in Fujitsu's acquisition history behind a parked domain. Meanwhile the law is in every undergraduate syllabus on the planet, and it gets cited harder every year that GPUs and many-core designs make people optimistic again.

There is a tidy irony in the shape of it, since the company's whole strategy was a bet on serial compatibility with somebody else's machine and the law is about the limits of doing things in parallel. Gene Amdahl died in November 2015 at ninety-two. He spent a career being right about limits.

@}-->--

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

« previous  ·  random  ·  next »