Alliant Computer Systems

~ the obit ~

Ron Gruner and Craig Mundie came out of Data General and started a company in Littleton, Massachusetts in May 1982, along with Rich McAndrew. They called it Dataflow Systems at first. The gap they were aiming at was real and enormous: a heavily configured VAX-11/780 did roughly one MIPS and one megaflop for about a million dollars, and a Cray-1S cost more than ten million. Nearly every engineer and physicist in the country was stuck on the cheap end of that gap, running simulations that took overnight when they should have taken an hour.

The answer announced in 1985 was the FX series, built out of 18 inch square boards that slotted into an active backplane holding a crossbar switch. The processors were called Computational Elements. Their scalar instruction set was Motorola 68000, but Alliant bolted on custom coprocessor instruction sets of its own for floating point and for vectors, and then a third one for concurrency. That third one is the reason the company mattered. A special bus between the computational elements handled synchronization at low enough latency that a compiler could take a single Fortran loop and hand out its iterations to all eight processors, with the hardware keeping them in step. The whole machine ran a version of 4.2 BSD called Concentrix.

Sizes were named for how many computational elements they held. An FX/1 was about as big as a full height PC and did about 2.5 MIPS against the VAX's one. A fully loaded FX/8 was air cooled and smaller than a VAX-11/780, roughly the footprint of a large photocopier. On problems that parallelized well it ran about five times an FX/1 rather than eight, which is Amdahl's law showing up on an invoice. In early 1988 the Advanced Computational Element replaced the Weitek floating point parts with a Bipolar Integrated Technology chipset and cut the vector cycle time to 42 nanoseconds. Existing FX/4 and FX/8 boxes could be upgraded in the field by swapping boards and updating a microcode file, which customers loved. That July, Alliant bought Raster Technologies for its graphics hardware and folded it into the VFX.

The trouble arrived from underneath. In 1990 the FX/2800 threw out the custom computational element entirely and moved to Intel's i860, a superscalar RISC chip anybody could buy. The CAMPUS/800 was announced in 1991, a cluster machine reaching 4.7 gigaflops on 192 nodes. By then the money was gone. On 21 October 1991 Alliant pulled its common stock and its convertible debentures off NASDAQ, saying plainly that it could no longer meet the capital and surplus requirements, and trading moved to the pink sheets. Gruner had already left. Mundie was running the company when it went bankrupt in 1992, with more than 650 machines installed at over 400 sites.

~ the rap sheet ~

~ the parallel loop ~

What Alliant did and what Convex did down in Richardson get flattened together by the shelf they shared, and they were not the same problem. Convex sold a vector machine with a vectorizing compiler, and the compiler's job was to find operations that could be done on many numbers at once inside one processor. Alliant's compiler had the harder problem. It had to prove that the iterations of a loop did not depend on each other, then hand them out across eight separate processors and collect the answers back in order. Alliant gave it instructions and a dedicated bus to do that with. Getting parallelism out of a machine without making the customer restructure twenty year old Fortran was the entire commercial proposition, and Alliant was good enough at it to end up contributing to the High Performance Fortran effort.

~ where the founders went ~

Craig Mundie started at Microsoft in 1992, the year Alliant folded, and set up the consumer platforms division, which is where Windows CE came from along with the handheld and automotive versions of Windows. He oversaw the WebTV purchase in 1997. When Bill Gates stepped back in 2006 his responsibilities were split between Ray Ozzie and Mundie, who took the research and strategy half and held it until 2012. He retired in 2014.

That is the half of this story that got anywhere near me. An Alliant went to a national laboratory or an engineering faculty and there were never more than a few hundred of them anywhere on earth; Pocket PC handhelds turned up in Turkish shops and I knew what they were. Gruner left before the end and founded Shareholder.com, an investor relations business that Nasdaq bought in 2006, which is a considerably softer landing than the one his first company got. The Computer History Museum has an FX/8 and an FX/1, and it acquired the FX/1 from Convex, the competitor that outlived them.

~ the killer micros ~

The phrase was coined at Lawrence Livermore and it described exactly what happened here. Through the late 1980s the workstation chips improved on a schedule set by volumes Alliant could not dream of, and every year the argument for buying a specialized parallel box at a specialized price got weaker. Gruner, asked about it long afterward, took the blame himself and named a disastrous relationship with the company's Japanese distributor along with a major product cycle it missed. He also pointed out that the performance and the price of ordinary workstations blew this whole market apart regardless of who was running which company.

The engineering won anyway, somewhere else. Symmetric multiprocessing with a shared cache and hardware synchronization is not an exotic arrangement now, it is the phone in your pocket, and the descendant of the automatic loop distribution Alliant sold hardware to support is a compiler flag that every working programmer has typed without thinking about it. Alliant needed customers to pay six figures for that arrangement in 1987. It ships for free in the toolchain now, and the company that proved it was worth having lasted ten years.

@}-->--

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

« previous  ·  random  ·  next »