Celerity Computing

~ the last words ~

In November 1984 a RISC processor was still mostly a stack of papers. Berkeley had published its work, IBM's 801 was an internal research machine nobody outside could buy, and Sun was selling Unix boxes built around Motorola parts. That month a company in San Diego started shipping a commercial machine whose entire architecture was RISC, and hardly anyone remembers that it did.

Celerity Computing was founded in May 1983. Steve Vallender and Nick Aneshansley left NCR to start it, along with Andrew McCroklin, who would still be working on what came out of it more than a decade later. The architecture was called ACCEL and it was built on the NCR/32, a chipset NCR had introduced in late 1982 and sold as components to anyone who wanted a 32-bit machine without designing a processor first.

The first product was the C1200 Personal Workstation, with a color display that went to 1280 by 1024. The C1230 and the C1260 followed, the second of those with a dual processor option. The operating system was 4.2BSD with System V Release 2 behavior merged into it, and by 1986 there was a symmetric multiprocessing version, which very few Unix vendors could say that year.

The machine that finished the company was the Celerity 6000, a minisupercomputer with a 33 MHz clock and room for eight processors in scalar or vector form. It was not done when the money ran out. Floating Point Systems, having its own bad year up in Beaverton, bought the assets and the technology. The trade press at the time dated the purchase to April 1988 and the record since has settled on September. Distressed sales close in stages and the paperwork runs behind the handshake.

~ the record ~

~ risc before risc sold ~

The argument in 1984 was that a processor could be made simpler and faster if you let the compiler absorb the difference. It was a good argument and it won, but it won from 1987 onward, when Sun put SPARC behind an entire product line and MIPS licensed cores to anyone with a fab. Shipping in 1984 meant selling into a market that had not yet agreed the idea was serious, to buyers who wanted to know what the next machine would be before they bought this one.

That was the harder question. A processor company needs a roadmap because customers are purchasing three machines forward. Sun had one and MIPS had one. Celerity had a chipset from a company whose revenue came from retail terminals and bank systems, and NCR had no reason to keep pushing that silicon because one customer in San Diego needed a successor part.

~ the chipset from a cash register company ~

The NCR/32 was serious engineering rather than a sideline. It arrived in the third quarter of 1982 as a set of parts, a processor chip with address translation and extended arithmetic and system interface pieces around it, and the unusual thing about it was that the microcode was external and writable. NCR put it in its own 9300 and 9800 machines and in a communications processor. Honeywell signed an agreement to look at it. Celerity took it somewhere NCR had no intention of going.

NCR arrived in my life as a name printed on things that take money. Supermarket tills, and the bank machines bolted into the wall outside branches, all through the 1990s in Turkey. It was years before I learned the same company had sold a 32-bit processor as a bag of components, and longer still before I learned somebody had built a Unix supercomputer out of it. A retail brand can stand in front of the engineering one for a very long time.

~ what shipped under the other badge ~

Floating Point Systems finished the Celerity 6000 and launched it as the FPS Model 500 in late 1988. One to four ECL RISC processors, 33 MIPS at the bottom and 133 MIPS with 167 MFLOPS at the top, priced from about £200,000 up to roughly a million. It ran Celerity's Unix, the 1986 hybrid. FPS renamed itself FPS Computing, stopped development on its Transputer-based T Series and told the trade press plainly that it had failed to take notice of what the market wanted.

The design never got another fresh start, but it did keep moving. FPS filed Chapter 11 in October 1991 and Cray Research took the assets that December. Cray sold the line to Silicon Graphics in 1996 and Sun bought it from there inside the same year, which is how work started in a San Diego startup in 1983 ended up in Sun's enterprise server division. McCroklin went along with every one of those handovers and was still there at the end of the chain.

@}-->--

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

« previous  ·  random  ·  next »