Attack of the Killer Micros versus Hairy Smoking Golfballs
Tony Viola writes:
"IBM Builds Linux Supercomputer to Speed the Search for Oil
"IBM eServer System Enables WesternGeco to 'See' Beneath the Earth's Surface, Saving Time and Money.........
"......The cluster at WesternGeco is comprised of 256 eServer xSeries 330 systems, each powered by two 933 MHz Intel Pentium III processors.
"Who would have thought 10 years ago that supercomputers would be built out of "PC chips". And, by IBM, who used to love building everything bigger than a PC out of proprietary IBM chips. And Linux too."
I could give the glib answer: this is a trivial consequence of Moore's Law. But I won't, as the actual issues are much more interesting than just Yet Another Invocation of Moore's Law.
Real Answer: Many, many people thought that supercomputers would be built out of "PC chips," and much longer ago than 10 years ago. Too much for me to write about in complete detail, but here are some items:
* I first saw a description on "hypercubes" using 8080As in around 1975. A presentation by IMSAI, which was developing a hybercube array to compete with Grumman's hypercube array (also built with 8080s). IMSAI later went to to fame with their IMSAI 8080, the first relatively solid personal computer (Altair 8080 was earlier, but not very solid). Intel used a bunch of IMSAI 8080s to control test equipment, circa 1976-79.
* By 1980-82, the Caltech "Cosmic Cube" had been designed and built, using another hypercube topology. A few years later, building on this work, Intel released the "Intel Personal Supercomputer," the iPSC. This came out in several versions, ranging from cabinet-sized small systems to room-filling very large systems.
* Several other companies had similar programs, including an Intel spin-off called Ncube. Teradata, I think it was, also sold arrays of Intel chips into places like K-Mart for inventory tracking. By the late 80s, there must have been a dozen competing architectures. MIMD and SIMD machines, fine-grained and coarse-grained. The Connection Machine was quite famous at its time.
* There was a very influential paper in the late 80s from a leading supercomputer expert. Entitled "Attack of the Killer Micros," it correctly pointed out the many advantages of CMOS VLSI (and ULSI) processors over conventional bipolar and ECL CPUS and predicted that large arrays of CMOS processors based on industry-standard architectures (Intel, MIPs, SPARC, PowerPC, etc.) would come to dominate the supercomputer market. This has of course happened.
* Not only are the world's largest supercomputers _all_ based on arrays of fairly standard CPUs (PowerPCs and Intel chips, mostly, but also SPARC and MIPS), but even smaller university supercomputers are mostly arrays of CMOS chips. The "Beowulf clusters" in use at Berkeley and most other universities, for example. (A group I co-founded, the pro-crypto group called "Cypherpunks" (a pun) was involved in this. Two UC Berkeley students, Ian Goldberg and Dave Wagner, used a Beowulf cluster of Intel chips to break an important cipher. This has happened several times, with various cipher challenges.)
* The "Deep Blue" chess computer from IBM was of course an array of such CPUs (some of them were Intel-based, others were PowerPC-based).
* The largest servers from Sun are based on arrays of processors. (Sun acquired the Oregon operations of Cray Research, which was using arrays of Sparcs. This Oregon operation was itself a spin-off or descendant or renaming (I'm not sure which) of Floating Point Systems, famed maker of the FP-120B floating point coprocessor box for the VAX. And that design was based on the work of some folks I knew at UC Santa Barbara in the early 1970s. A small world.). This server market for Sun is quite successful, accounting for billions a year in revenues.
* Hypercubes and Beowulf cluster arrrays, in addition to the 64-way SMP servers made up of Sparcs and other CPUs, are dominant.
In summary, it was very obvious to me and to many others that the future of computers lay in very large arrays of mass-produced processors. I'd say this was a foregone conclusion by 1985, if not as early as 1980. (A friend of mine in Oregon had worked on the HEP, Heterogenous Element Processor, for Denelcor, and he filled in some of the reasons for me as early as 1980. Another friend of mine worked for Masspar a few years later and he confirmed the same points. And Intel's Personal Supercomputer was, as I said, making headway in various large national labs. Intel had the fastest computer in the world for a year or two....until an even faster array of inexpensive processors was built....the generations continue to evolve rapidly. I expect an array of P4s may take the speed crown again soon.)
The last gasps of convention "hairy smoking golfballs" (explained below) were probably these efforts:
-- the very hard to cool processors of the IBM mainframes...IBM had to develop the sophisticated "thermoconduction modules" with plungers and whatnot to cool the CPUs of their bipolar machines
-- The Cray 2 was a non-parallel (in the non massive sense...it had parallel vector elements, of course) machine which was in many ways Seymour Cray's last delivered machine. His Cray 3 was also not a parallel machine, and perhaps you have heard of his very arduous efforts to solve the packaging problems with his GaAs processing elements (made by GigaBit Logic, if I am remembering correctly). Just before his death, he had set up another company and was planning to use--drum roll!--an array of Intel processors!
-- what I think of as the "ultimate last gasp" for bipolar or ECL processors was Trilogy. Doug Peltzer set out to built a very high speed mainframe using bipolar and ECL technologies. They failed, and Trilogy folded.
So, I hope this answers the "who'd've have thought" point!
(No insult intended for Tony...this is just a good excuse for me to write about the long history of massive arrays of cheap processors.)
--Tim May
(Supercomputer gurus refer to such designs as "hairy smoking golfballs": the CPU must be not much bigger than a golfball, else the cycle times are just not short enough because of speed of light issues, the processor will have thousands or tens of thousands of wires coming off of it, and the kilowatts of bipolar or ECL power dissipation will be extraordinarily hard to dissipate, even with massive water flows. Trilogy had to deal with all of these issues, while CMOS VLSI was contantly getting faster and faster.) |
