MasPar

~ the eulogy ~

Jeff Kalb was running the division at Digital Equipment that made integrated circuits when the people around him worked up a machine architecture descended from the Goodyear MPP, an older array processor built for NASA. He left in 1987, set up in Sunnyvale, and built it. The shape of the thing is easy to describe and was hard to sell: one control unit reads one instruction, and thousands of small processing elements all do that same instruction at the same moment to their own piece of the data. Nothing branches independently. If your problem is an image or a large numerical grid, this is close to free performance. If it is anything else, you own an expensive room heater.

The MP-1 shipped in 1990 and the MP-2 followed in 1992, with the later processor chip carrying 32 elements that were each a full 32-bit ALU. Over 200 systems went out, which put MasPar second in its category behind Thinking Machines and well ahead of the rest of it. Digital resold the hardware as the DECmpp, so a fair share of the installed base never carried the maker's name on the front panel at all. Then the category went. Thinking Machines filed for bankruptcy in 1994, the federal research money that had been buying these machines went to clusters of ordinary workstations, and the SIMD argument was over before anyone finished making it. There was no MP-3.

In June 1996 MasPar stopped hardware development entirely. I was twelve years old that month, in Turkey, about two years away from my first modem and considerably further from any machine like this. The age is the boring half of that distance. The sharper half is that if a Turkish institution had somehow bought one, the crate would have said Digital on the side, because a nine-year-old company in Sunnyvale had no route to a buyer that far out except through somebody else's sales force.

~ the dossier ~

~ sixteen thousand of them ~

A MasPar was not a computer on its own. The processor array sat behind a front-end workstation that ran the operating system and handed the array its work, which is how most attached parallel hardware of the period was arranged. What you bought was a very wide piece of arithmetic with a host bolted to the side of it.

The economics of that were unusual while they lasted. Because every element ran the same instruction, the control logic was paid for once instead of 16,384 times, so the per-processor cost was extremely low compared to a machine where every CPU had to think for itself. That is also the whole vulnerability. The moment general-purpose microprocessors got cheap enough to buy by the hundred and wire together, the customer could have independent branching for about the same money. Thinking Machines had already conceded the point with the CM-5 in 1991, which abandoned lockstep for ordinary SPARC processors, and it did not save them either.

~ the software company underneath ~

What MasPar did next is the part that reads oddly from 2026. Instead of liquidating it kept the staff and took $6 million in equity from Perot Systems along with Kleiner Perkins and two other backers, relaunching in June 1996 as NeoVista Software with a data mining line called the Decision Series. The logic held together. The hardware had always been built to sweep enormous flat data sets, and retailers were just then starting to want exactly that run against their transaction logs. Hewlett-Packard and Informix signed on as partners alongside Perot.

NeoVista lasted three more years on its own and was bought by Accrue Software in 1999. Accrue sold the division to JDA Software in 2001, where the retail forecasting work carried on under yet another name. The processors are what people remember and the company's last act was to admit they were the disposable half.

@}-->--

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

« previous  ·  random  ·  next »