Kendall Square Research

~ the last words ~

Kendall Square Research was founded in 1986 in Kendall Square in Cambridge, a few streets from MIT, by Steven Frank and Henry Burkhardt III. Burkhardt's record was about as good as the industry had to offer. He had been on the original PDP-8 design team at Digital and went on to co-found Data General. Money followed the resume. The company moved out to Waltham as it scaled toward production, chasing a machine that would make massive parallelism easy to program.

The first KSR1 went in during 1991. The processor was the company's own 64-bit RISC design running at 20 MHz for 40 megaflops, and anywhere from eight to 1,088 of them sat in a hierarchy of rings. Underneath was a cache-only memory architecture. There was no main memory in the ordinary sense and no home location for any page. All 32 megabytes per node were cache, and data migrated and replicated toward whichever processor was touching it, coherently, in hardware. The company sold this as ALLCACHE. Customers watching early systems fall over called it Allcrash.

Reliability was one problem and arithmetic was the other. A KSR1 processor in 1991 ran at 20 MHz and 40 megaflops while an Intel 486 sitting in a shop window ran at 50 MHz and 50 megaflops. When DEC shipped the Alpha in 1992 at up to 192 MHz and 192 megaflops, the KSR2 came back with 40 MHz and 80. A small company had taken on processor design and system design at once because no 64-bit part existed to buy, and by the time its machines shipped, one did.

I was ten years old in turkey when this collapsed and there is no personal angle available here. A machine priced above a building, sold to national laboratories, is not something a household ever had a relationship with. What has followed me since is the pitch rather than the product. One flat address space, with memory that moves to whichever processor is using it, so nobody writing code has to know which node the data physically sits on. Every generation of parallel hardware since has offered a version of that promise and every generation of programmer has eventually found the seam in it. Kendall Square went bankrupt building the idea in 1991 and the seam is still there in 2026, on hardware I actually use.

~ the record ~

~ allcache ~

The architecture was not a gimmick and the numbers say so. Traffic between nodes ran at up to 4 gigabytes per second. A full 1,088-node KSR1 could carry 510 input and output channels aggregating over 15 gigabytes per second, which was a serious figure in 1991. The processor was a two-wide VLIW with a branch instruction that let the compiler decide whether to keep the results of the next two instructions depending on how the test resolved. Delay slots were not interlocked, so the compiler had to schedule around them or the machine loaded the wrong values.

The company also went after the commercial parallel database market alongside Oracle rather than staying inside scientific computing. That was a fair read of where the money was. It also meant a small firm was designing its own processor and porting an operating system onto it while trying to sell databases to businesses, which is three companies' worth of work funded by one company's revenue.

~ the fall ~

In october 1993 Kendall Square postponed its third-quarter results, explaining that auditors were examining revenue figures for signs that sales had been recorded in error. That is the sentence that ends companies. On Monday 29 november Price Waterhouse withdrew its report verifying the 1992 financial statement and warned of substantial downward adjustments. The stock fell 33 percent in a day, $3.75 down to $7.50. It had touched a record $25.75 on 3 september.

Two days later came the management shakeup and the actual number. Revenues for 1992 and 1993 had been overstated by more than $18 million. The 1992 figure went from $20.5 million to $16.3 million and the 1992 loss went from $12.7 million to $17.2 million. Revenue for the first nine months of 1993 came in at $10.6 million, some $14 million below what the company had already reported for the first six. The board announced stringent new controls on revenue recognition, which is the thing a board announces afterwards.

The mechanism was mundane and specific to this business. Universities and national laboratories order supercomputers a year or more before delivery, so when an order becomes revenue is a genuinely arguable question, and Kendall Square answered it early on machines that had not been paid for. Burkhardt had sold 50,000 shares that august, thirteen percent of his holding, weeks before the peak. Shareholders sued in federal court in Boston and the SEC was already reading.

~ what it became ~

The dates after that are contested and the contest tells you something. Wikipedia says the company went out of business in february 1994. The SEC's own filing says the stock traded on NASDAQ until april 1994, when the NASD delisted it. Koch, who by then had put in $65 million and held 29 percent, ended production in september 1994 after the KSR2 failed to arrive as a reliable machine. Bankruptcy came in january 1995 and the company was liquidated. Take whichever of those you like as the death, because it took a year of them.

The SEC filed in april 1996 against the company and four of its officers, over six quarters of materially false statements across 1992 and 1993 and over stock sold while holding information the public did not have. Nobody admitted or denied anything and everybody settled. Burkhardt paid around $1.1 million. Two others were ordered to pay sums in the hundreds of thousands and had nearly all of it waived down to $40,000 each on a demonstrated inability to pay more, having already spent the proceeds.

One customer refused the ending. Pacific Northwest National Laboratory, a Department of Energy site, had bought KSR2 hardware and reacted to the collapse by purchasing an enormous stock of spare parts, then kept its machines running for years after the company that built them stopped existing. Somewhere in that is the fairest verdict available. The machine was good enough that its owners hoarded parts for it, and the accounting was bad enough that there was nobody left to call.

@}-->--

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

« previous  ·  random  ·  next »