Afara Websystems

~ the eulogy ~

Kunle Olukotun was a Stanford professor who kept hearing the same complaint from people who ran data centers. They were not running out of processor speed. They were running out of electricity and floor space. Around 2000 he decided that the answer was to stop chasing clock rate and build a server chip out of many deliberately simple cores, each one handling several threads, so that whenever one thread stalled waiting for memory another could use the pipeline.

He recruited Les Kohn, a microprocessor architect who had designed chips at both Sun and Intel, to do one more. Fermi Wang came in as the third founder and acting chief executive. They set up in Sunnyvale and called the company Afara Websystems. Olukotun later explained the name to IEEE Spectrum: afara means bridge in Yoruba.

The pitch to venture capitalists was that the design could improve server performance roughly tenfold over what Intel was shipping. Almost nobody believed it. Sequoia Capital and Raza Foundries did, and the company got far enough to run instructions through a software model of its pipeline inside three months of starting the design. The plan was to sell whole servers rather than chips, built at TSMC and running Linux.

~ the dossier ~

~ eight cores ~

The architecture had eight cores. Each had a short, simple pipeline and juggled four threads, which gave thirty two threads on one piece of silicon at a time when a high-end server processor had one or two. Because each core was simple, the whole thing could run at a lower clock and draw less power than its competitors, and because the threads covered each other's memory stalls, the pipelines stayed busy.

In 2000 this read as eccentric. Every benchmark that mattered commercially measured how fast one thread finished, and on that test Afara's chip lost every time. What it would win was work per watt and work per rack, which is a thing you only care about when you are paying the power bill. My own work now is computer vision on hardware bolted to a drone, where the power budget decides everything and single-thread speed is close to irrelevant. From that seat the Afara position reads as plainly correct, and it takes some effort to remember why it was a hard sell in 2000. It was hard because the people buying servers then were not yet the people paying for the electricity to run them.

~ the motel ~

Funding dried up during 2001. The founders were in New York on 11 September 2001 to meet potential backers, and spent the next two days stuck in a motel before that meeting finally happened. After that autumn the venture market stopped writing checks for expensive, long-horizon silicon, and Afara was proposing to design a server processor from scratch. It never raised again.

~ what shipped without them ~

Sun's quarterly filing for the period ending 29 September 2002 handles the acquisition in a few cold sentences. "During July 2002, we acquired Afara Websystems, Inc. (Afara), a development stage enterprise." The total purchase price was approximately 28 million dollars, about 27 million of it in roughly 5 million Sun shares issued or issuable, plus a million in transaction costs. On 22 July 2002 Sun issued 4,867,000 shares and 109,000 warrants to Afara's stockholders in exchange for all of their ownership interests.

The purchase price allocation is the part worth reading twice. Sun put 68 million dollars of the value on the assembled workforce and 4 million on unearned stock compensation, then subtracted 17 million of net liabilities assumed and 27 million of deferred tax, which is how the total lands on 28 million. There was no in-process research and development charge at all, and the deal was accounted for "as a purchase of a group of net assets rather than a business combination". Sun's accountants looked at Afara and concluded, in the technical language available to them, that the company was its engineers.

Those engineers became the Niagara project, which grew to about a hundred people inside Sun. The design shipped on 14 November 2005 as the UltraSPARC T1, sold with up to eight cores and four threads on each of them, thirty two threads in all. That was three years and almost four months past the end of the company that conceived it. Sun released the T1 design under the GPL as OpenSPARC on 21 March 2006, which made it the rare commercial processor anyone could read. The T2 doubled the threads per core and the T3 reached sixteen cores and 128 threads, and the whole line ran on into Oracle's SPARC servers. David Yen, the Sun executive who took the deal, said at the time that Afara brought Sun "a seasoned and talented team of SPARC chip designers". The filing had already said the same thing in accountant's language, and both of them turned out to be literally true.

@}-->--

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

« previous  ·  random  ·  next »