Denelcor

~ the obit ~

Six. That is how many Heterogeneous Element Processors Denelcor delivered in total, between 1981 and 1985, and the last of them shipped the year I turned one. I am not going to claim much distance from this one, because there is barely any distance to claim: the complete population of human beings who ever put their hands on a HEP would fit in a lecture room, and every one of them worked somewhere you could name off a list.

David R. Miller founded Denelcor in Denver in January 1968. In 1973 NASA came asking for a parallel supercomputer and the company started designing one, with Burton Smith as vice president of research and development and the principal architect of what came out. Smith was solving a problem that everybody had and almost nobody was attacking head on. Memory is slow compared to the processor and the gap widens every year. Most machine designs of the period handled that by adding caches and hoping the program had an access pattern the cache liked.

The HEP answer was to stop waiting. Each processor module kept a queue of up to fifty program status words, one per live process, and issued one instruction per cycle from a different process each time through an eight-stage pipeline. By the time a given process came back around for its next instruction, its previous one had finished. No instruction in the pipeline ever depended on the one in front of it, so the interlock and bypass logic that makes a fast pipeline complicated was simply not needed. If a process stalled on memory, the machine ran the other forty-nine. A single process went no faster than it would have anywhere else; the throughput came from there being fifty of them.

Synchronization got the same treatment. Every register and every location in data memory carried a full or empty state alongside its value, and a read could be told to wait until the location was full or a write to wait until it was empty, which put the producer and consumer handshake into the memory hardware rather than into a library. Programmers marked the variables that needed it with a dollar sign. This is a genuinely small amount of mechanism for what it buys, and it is the part of the design people still write papers about.

The business side never came together. The machines were expensive and the largest configuration ever delivered had four processor modules against a theoretical sixteen, so the price per unit of delivered performance stayed stubbornly high. Buyers came down to a short list of institutions with parallel programming research budgets. Denelcor closed in 1985.

~ the rap sheet ~

~ six customers ~

The delivered systems went to the National Security Agency, to Los Alamos and Argonne national laboratories, to the Army's Ballistic Research Laboratory at Aberdeen Proving Ground in Maryland, and to Messerschmitt in Germany. Aberdeen had the four-module machine and had wanted it for fluid dynamics. Messerschmitt appears to have been the only site running the thing on real production work; the rest used it to experiment with parallel algorithms, which was valuable to computer science and produced a decade of literature. Design work had started in 1973, so twelve years of engineering went into six sales.

~ the idea that outlived it ~

Burton Smith spent 1985 to 1988 as a fellow at the Supercomputing Research Center, then co-founded Tera Computer Company to build the same idea again with better silicon. Tera's MTA kept the fine-grained multithreading and kept the full and empty bits. In 2000 Tera bought the Cray Research name from Silicon Graphics and renamed itself Cray Inc, which means the Denelcor design philosophy ended up wearing the most famous badge in the industry. Smith went to Microsoft as a technical fellow in 2005 and died in 2018.

The approach won completely, just nowhere near supercomputing. Every GPU on sale today hides memory latency exactly the way the HEP did, by keeping enormous numbers of threads resident and switching among them every cycle so the memory system never stalls the arithmetic; Intel shipped a two-thread version of the same trick to desktops in 2002 and called it hyper-threading. Denelcor sold six machines and lost the company proving an idea that now runs in every laptop and every phone, under other names, for buyers who were not born when the last HEP left Denver.

@}-->--

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

« previous  ·  random  ·  next »