Floating Point Systems

~ the last words ~

Norm Winningstad was a Tektronix engineer before he started Floating Point Systems in Beaverton, Oregon in 1970 with three partners. The product idea was narrow and it worked for twenty years: most scientific programs spend nearly all their time in a handful of numerical loops, so rather than sell a whole fast computer, sell a box that does only those loops and let the customer keep the computer they already own. The AP-120B came out in 1976 and attached to a minicomputer as a peripheral. The host handled the files and the terminals; the box did the arithmetic, at speeds the host could not approach.

Oil companies bought them by the crate for reflection seismology, where the whole business is running enormous Fourier transforms over field data. Physical chemists were on the customer list, and so were the NSA's cryptologists. The FPS-164 in 1981 moved to 64-bit arithmetic and attached to IBM mainframes and VAXes, and the FPS-264 that followed swapped TTL logic for ECL and ran about five times faster. Hundreds of these boxes shipped, possibly thousands, and for a while Floating Point Systems was the most successful company in a category it had more or less invented.

Then it tried to become a computer company. The T-Series in 1986 was a parallel machine rather than an attachment, and it was a commercial disaster that took years of engineering money with it. The company pivoted again to SPARC-based file servers with the Model 500 line, which was a sound enough idea and arrived too late to refill the accounts. Floating Point Systems filed for reorganization under Chapter 11 in late October 1991. Cray Research bought the superserver assets that December for $3.25 million.

~ the record ~

~ the attached processor ~

The attached array processor is a category that has almost no shelf presence in computing history and was, for about a decade, an obvious purchase. If you had a VAX and your problem was numerical, a box on the side of it bought you performance that a whole new machine could not, without rewriting your operating environment or retraining anyone.

The catch was the software. Speed came from calling FPS's hand-written math library rather than from recompiling, which married your code to one vendor's hardware for as long as you wanted it to stay fast. Then the gap that justified the whole arrangement closed. Intel put a usable floating point unit on the 486 die in 1989, and a peripheral whose only selling point was arithmetic had nothing left to sell.

~ the t-series ~

The T-Series deserves its own line because of what it was built from. Each node paired an Inmos transputer with a Weitek vector unit, wired into a hypercube, and the transputer's job was communication while the Weitek did the arithmetic. On paper it scaled beautifully. In practice almost nobody could get real programs running well on it, and the largest configurations were mostly demonstrations. The 1986 launch became the expensive middle of the company's life.

~ chapter eleven then the sale ~

Three and a quarter million dollars is an unusually small number for a twenty-one-year-old company with real products in the field. It is a bankruptcy price, which means it reflects what a buyer could get away with paying rather than what the engineering was worth, and Cray Research was the only serious bidder in the room.

What Cray got was the Model 500 line and the people who built it. That became Cray Research Superservers, Inc., later folded in as the Cray Business Systems Division, and the Model 500EA was relaunched in 1992 as the Cray S-MP. The CS6400 followed in 1993, a SPARC machine with up to 64 processors, sold to customers who wanted large symmetric multiprocessing without a Cray-sized budget.

~ what sun got ~

Silicon Graphics bought Cray Research in 1996 and immediately had a problem: the business systems division sold SPARC servers, and SGI sold MIPS. It put the unit up for sale and Sun Microsystems bought it, at a price the public record only pins down loosely, somewhere between a reported $17 million and something well under $100 million.

Sun turned the CS6400 design into the Ultra Enterprise 10000, the machine everyone called Starfire, and sold roughly 7,000 of them into the dot com build-out at margins that ran into billions. The design lineage goes back through Cray Research Superservers to the Model 500, to a bankruptcy in Beaverton, to four men who left Tektronix in 1970 to sell a box that did nothing but arithmetic. Somebody paid three and a quarter million for that and somebody else collected.

@}-->--

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

« previous  ·  random  ·  next »