Montalvo Systems

~ the goodbye ~

The company began in 2000 as MemoryLogix, in the garage of Peter Song, who had previously written for Microprocessor Report. It announced a licensable x86 core at the Microprocessor Forum in 2002 and never shipped it. In 2004 Vinod Dham pulled Song's small group together with two more veteran x86 architects, Greg Favor and Mike Yamamura, and in 2005 the result was renamed Montalvo Systems and handed to Matt Perry, who had just finished being chief executive of Transmeta.

The plan was a low-power x86 processor for portable machines, originally with asymmetric cores: a couple of fast ones for real work and slower, cheaper ones for everything else, which is roughly the arrangement every phone and laptop uses now. That design was abandoned in 2007 in favor of four identical cores, on 65 nanometers, after the company decided performance was what it was actually being judged on. It taped out its first chip around 13 February 2008.

There is no honest personal angle here, because no such chip was ever in anybody's hands. Montalvo never sampled a part to anyone, let alone sold one, and Microprocessor Report could not even find an issued US patent in its name. The entire company is a thing you can only read about, and the only reason it is readable at all is that the trade press of spring 2008 happened to be paying close attention while it fell apart.

~ the chronicle ~

~ named after a mansion ~

Villa Montalvo is an estate in the hills above Saratoga, and in 2000 it was where Transmeta chose to come out of stealth mode and announce the Crusoe processor. Peter Glaskowsky, a former editor-in-chief of Microprocessor Report who went on to join the company, suggested naming the new venture after it. Tom Halfhill, writing the obituary in the same publication in May 2008, allowed himself one line about that, wondering in print whether the new name had been a bad omen.

The Transmeta lineage ran deeper than the name. The Register described the company in February 2008 as full of Transmeta veterans, and the chief executive had run Transmeta. Halfhill's comparison is the one worth keeping, because it is not flattering to either: Transmeta at least shipped a few million processors and eventually extracted a cash settlement from Intel. Montalvo managed neither.

~ the verification problem ~

The design was not the problem. Proving that the design behaved like an x86 was the problem. By the end, according to the post-mortem reporting, at least half the engineers were employed characterizing how existing Intel and AMD chips actually behaved, and building a test suite meant to cover every permutation of software written in the previous twenty-five years. Halfhill had just toured Intel's verification labs in Austin, two floors of heavily staffed equipment devoted to one new processor, and made the obvious point that a startup cannot reproduce that.

The architecture had also been moving away from them the whole time. x86 had picked up 64-bit integer processing in the preceding years, along with extensions for virtualization and for wide media instructions, and each of those is another surface a clean-room clone has to match exactly. Burn rate was about three million dollars a month. Finishing tape-out needed something like forty million more, and reaching volume manufacturing something nearer a hundred million, which the company was trying and failing to raise through the spring of 2008. Halfhill's summary is the sentence the whole thing reduces to: the company evidently ran out of money before the engineers ran out of problems.

Intel shipped Atom that same year, into precisely the low-power x86 niche Montalvo had been aiming at for three years. Microprocessor Report covered Atom on 7 April 2008. Fourteen days later Sun bought the remains.

~ what sun took home ~

On 31 March 2008 Montalvo laid off two-thirds of its staff. In mid-April it disbanded the rest, nearly three hundred people, and stopped all work on the unfinished processor. Sun Microsystems completed its purchase of what was left on 21 April 2008 and said nothing publicly until the 24th, at which point a spokesperson confirmed the earlier date and declined to give a price on the grounds that the acquisition was not material to Sun's earnings per share. Press estimates of that price ranged from a little over five million dollars to north of ten, which is a factor of two and tells you how little anybody actually knew.

Sun hired more than a hundred of the engineers, Perry included, and folded them into its microelectronics group. No Sun or Oracle product ever carried the Montalvo name, and no announcement was ever made cancelling the chip, because there was nothing to cancel that had been announced in the first place. The venture lender's filings are the cleanest record of the ending: Montalvo Systems sits in the loan schedule for the quarter to 31 March 2008 at $453,531 outstanding, and in the schedule for the quarter to 30 June it is simply absent.

@}-->--

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

« previous  ·  random  ·  next »