The Following appeared as Article 9037 on HPC wire on 27th March 1995 recording the end of the Cray-3 project and Cray Computer Corporation.
CASH-STARVED CRAY COMPUTER CLOSES, SEEKS CHAPTER 11 Mar. 27 NEWS FLASH by Norris Parker Smith, Editor at Large HPCwire =============================================================================
Why did Seymour Cray's company, Cray Computer Corporation (CCC), descend into bankruptcy despite his unmatched contributions to computing, his remarkable skills at computer architecture and manufacturing, and his exceptional persuasiveness? How will the absence of CCC affect the high-performance market -- and the high-performance computing community? The answer to the first question consists of a few simple propositions: o The world changed; o Seymour Cray tried hard to keep up with those changes without compromising his own vision of high-performance computing; o Toward the end, he tried very hard indeed. Adjusting prices downward and yielding to the demands of modern marketing, he accepted that the grace, ingenuity and redoubtable performance of his solutions might not be enough to close sales; o Despite these efforts, he could not obtain the customers he needed so desperately; o If solid revenue had been in reach, he might have been able to raise more funds; o Neither hope came true in time. Just a few weeks ago, stockholders approved an additional stock issue. Placement was attempted outside the United States in order to avoid the delay required for SEC filing. The foreign investors evidently decided that the risk was too great. Lack of cash forced CCC to close and most of CCC's 350 workers were dismissed. THE IRONY OF ORDINARINESS Two clauses in these statements require further examination: "The world changed" and "his own vision of high-performance computing." How has the world changed? Some commentaries on CCC's decision to seek the protection of Chapter 11 have attributed it to a decline of supercomputing brought about by the loss of easy defense money and the proliferation of killer microprocessors. This is mostly to the point, but supercomputing has not declined. It is, in fact, thriving to a degree never approached while Seymour Cray and supercomputer were identical in the public mind, like Einstein and relativity. Supercomputing is thriving because it has changed so much that the old term -- always half-description, half-slogan -- has acquired misleading implications. The days of expensive, delicate machines, presided over by specialists and used mostly by people with exceptional requirements, are passing. Supercomputing has become ordinary; just another way to do computing. In this perspective of his long and productive career, this outcome means that Seymour Cray's lifework has been a remarkable success. Ironically, as Mr. Cray defined the objectives of CCC and the character of its products, it was unable to find a place in a world of ordinariness. Seymour Cray has always accomplished the extraordinary and he wanted to do it one more time. GOLDEN YEARS In the golden years, supercomputing supported one profitable medium-sized company, Cray Research, three Japanese emulators that probably rarely made much genuine profit doing it, and a few small companies pursuing different paths, like Thinking Machines (TMC) and Intel's supercomputing (now scalable) computer division. Both TMC and Intel were dependent to some degree on outside funds -- DARPA subsidies for Thinking Machines and support for Intel's parallel-processing venture from the central treasury at Intel headquarters where the money is made. Dozens of would-be Craylets rushed to occupy a narrow niche called mini- supercomputers. Only Convex survived, doing well yet significantly smaller than Cray Research. Thus, there was only one influential, self-sustaining operation in the industry, Cray Research. It was shaped to a remarkable degree by Seymour Cray's personal creativity, persistence and skill. If one includes the earlier years at Control Data when Seymour Cray was beginning to establish his way to do heavy-duty computations, the span extends to almost two decades. In modern times, few individual creative persons have shaped so much of a significant technology for so long. FOUR CHANGES In recent years, however, supercomputing has changed in four ways: o A raid-the-junkyard philosophy has been adopted. Microprocessors, other components and, in some cases, whole modules are borrowed from workstations or other technologies -- thus reducing costs and making it easier to raise reliability. o It has become yet another corporate phenomenon. Big companies like IBM, Digital, AT&T and Hewlett-Packard are now playing important roles. Silicon Graphics, which started a number of years after Cray Research concentrating upon specialized graphics workstations, now has more than twice the revenues of Cray Research. It has, with considerable success, gone into the lower end of supercomputing -- almost as a sideline. o Sales are increasing greatly. At the lower end (prices under $2 million), sales in the scientific/technical market alone are expected to reach about $1.5 billion this year, 50 percent more than the highest figure reached during the golden age. For years, despite intermittent effort, supercomputing made few inroads into commercial markets -- perhaps three times as large as the scientific market. Although much skepticism remains among commercial customers, significant progress is now being made. o Ironically, sales are growing primarily in the lower levels, in systems with a few dozen -- or fewer -- processors and various strategies to share a central memory. In principle, Seymour Cray would be at home with this kind of architecture. Indeed, his efforts at CCC -- the CRAY-3 and CRAY-4 -- were in the most basic terms shared-memory multiprocessor systems with similar numbers of processors. The big difference nowadays: each multi-megaflop (soon to be gigaflop) processor of a typical shared-memory multiprocessor sells for a few tens of thousands of dollars, not a couple of million. In late 1994, CCC brought its price per GFLOPS for the CRAY-4 down toward the levels of its competitors, but the change came too late. ADAPTATION This is not simply a reenactment of that great drama of American nostalgia, the worthy individualist whose high-quality small business or delightful country store is squeezed out by faceless, homogenizing corporate giants. The troubles of CCC can be illuminated by considering Cray Research which retained the imprint of Mr. Cray for some years after he left it, and Seymour Cray's historical baseline customers: the code-breaking/intelligence agencies and the big federal laboratories. Since the late 1980s, Cray Research has been seeking to adapt to the realities of a changing world. It has been a protracted struggle marked by the extrusion of two of the main protagonists: Steve Chen and Seymour Cray himself. Mr. Cray left in 1989 to, in effect, set up his own country store, operated according to his distinctive principles. Cray Research has gone on to expand its product line to a degree that would have been high heresy in earlier decades. It has now embarked upon a reconsolidation based on a multi-layer configuration, also pursued by Convex and other makers, that may now become a prevailing solution, especially for large-scale problems and the ascent of Mt. Teraflop. Some new models even include memory caches, another heresy. Various euphemisms are applied by Cray Research and others, but "cluster" is the simplest description of the top layer of these proposed hierarchies, where nodes consisting of shared-memory systems are aggregated into a distributed supersystem. ANAGRAM OF LATENCY For a purist like Seymour Cray, cluster is just an anagram of latency. Bandwidths may be rising sharply and cleverness can disguise the effects of high latency, but it is inescapable. Seymour Cray has always preferred straightforward, simple solutions, combined with a willingness to invest much money and originality in the computational infrastructure that makes simplicity possible. His favorite mathematical statement is, in effect: "the best connection between two components is the shortest possible line." He rejected fancy ideas like caches that, in his view, introduced self-defeating complexities. Mr. Cray also emphasized fundamental virtues: plenty of bandwidth, especially to memory, careful attention to memory management, and a sound balance among processor power, memory and I/O. He quietly criticized, for example, parallel systems that claimed huge aggregate performance but were so imbalanced that actual capabilities were a small percentage of the claimed performance. HIGH OVERHEAD CCC was established in the hopes of achieving exceptional performance while pursuing these pure goals. In doing so, Mr. Cray largely avoided the borrow-parts-don't-make-them philosophy that now dominates most of high-performance computing. His strategy was to fit very fast processors into innovative, compact packages. The chips would be fast because they would be based on gallium arsenide (GaAs) rather than silicon. Mr. Cray chose, moreover, the most demanding of two basic ways to use GaAs. This led to much frustration and delay. Finally CCC acquired its own GaAs foundry. More time was spent achieving acceptable yields with this brittle, unfamiliar material. The intricate design and compact dimensions of the processor nodes required special light-fingered robots. Both were remarkable accomplishments, and Seymour Cray relished as always his ability to reach goals that others had scoffed at. Nevertheless, these accomplishments were also causes of added expense and delay. CCC thus ended up with exceptionally high manufacturing overhead while other vendors of high-performance systems were increasing modularity to keep costs down. Furthermore, many competitors were moving toward production rates in the hundreds of units per year while CCC's costs would be spread over a dozen or so units at best. None of these burdens would have been crippling if CCC's first product, the CRAY-3, had offered distinctively higher performance compared with preceding models and could sell into an unoccupied marketing window. Instead, Cray Research's C-90 reached the market at about the same time with similar performance. As CCC spokesmen later acknowledged, the timing was bad. CCC did make one tentative sale, to Lawrence Livermore National Laboratory (LLNL) for support of the Department of Energy's nationwide research program. In December 1991 when CCC was unable to meet delivery/performance goals, the order was cancelled. A small CRAY-3 was later placed at the National Center for Atmospheric Research (NCAR) where, after some time and effort, it became ready for production work. This was not followed, however, by an actual sale. DEDICATED, LOYAL CUSTOMERS From his early days at CDC, Seymour Cray depended upon a foundation of dedicated, loyal customers that could be counted upon to buy the single-digit serial numbers -- and thus provide cash flow for expenses required by later manufacturing and sales. This roster included national laboratories like LLNL, National Science Foundation-supported sites like NCAR and, later, the national supercomputing centers. Above all, Mr. Cray looked toward the cryptanalysis/image analysis/signal analysis requirements of federal intelligence agencies. Once again, these were the customers that Seymour Cray counted upon to put CCC on track toward customer acceptance and adequate cash flow. It never worked out. These agencies chose to stay with Cray Research which offered similar performance and much better prospects of financial stability. Even worse, these agencies became deeply interested in clustering -- of classical C90 vector supercomputers and, later, of Cray Research's new J90 scaled-down, modernized, shared-memory multiprocessors. Seymour Cray pressed ahead with the CRAY-4, an even-more-so machine with tighter design parameters and roughly twice the per-node performance of the CRAY-3. If CCC had not already spent all the money it could get, the CRAY-4 might have won some customers away from Cray Research's counterpart, the T90. Now, no one will ever know for sure. WIDER EFFECTS Most customers and competitors have undoubtedly anticipated -- and discounted -- the departure of CCC for some time. That would have been considered an inescapable outcome if it were not for the unique aura of Seymour Cray's name and personality. For most observers, timing was the only question. Now that uncertainty has ended, Cray Research may find it somewhat easier to sell T90s, due for volume shipment during the latter half of this year. The T90 will need all the luck it can get. The market for upper-end vector machines has shrunken, due to loss of market share to scalable parallel systems and to shared-memory devices like Cray's own J-series as well as decline in the supply of defense money. It would be a shame if CCC's technological achievements in taming GaAs and robotic manufacture would be allowed to gather dust. Seymour Cray himself suffers from his status (despite his own best efforts) as a mysterious celebrity as well as a highly-talented computer designer. This could be damaging: people have very short memories about celebrities and even shorter attention spans. Despite the unfortunate course of Cray Computer, however, Seymour Cray should -- and, it is to be hoped, will -- continue to receive respect for his persistence as well as his imagination and the contributions made during a long and unique career. Above all, his emphasis on the fundamentals on well-balanced systems deserves continuing reiteration. If Mr. Cray wishes to moderate somewhat his habit of silence, his comments could be valuable contributions to the ongoing dialogue of high-performance computing. Seymour Cray, like many quiet men, combines a gentle humor with a well-developed sense of irony. He was probably amused to hear that a hasty reporter in Minnesota misunderstood the initial news of CCC's filing for bankruptcy and told the world that Cray Research was seeking Chapter 11. This led to much fuss and a distinct outflow of resumes before the truth was re-established. Supercomputing may have become ordinary, yet the name Cray has not. It still commands respect -- and, at times, can cause confusion. ********************************************************************** HPCwire has released all copyright restrictions for this item. Please feel free to distribute this article to your friends and colleagues.