computing power was increased by several orders of magnitude during World War II, when federal interests in cryptanalysis and other wartime needs led to the development of the vacuum tube computer, and in the late 1950s, when federal interests in military command and control led to transistorized computers.4As economist Kenneth Flamm has observed, "By tracking the origins and history of key pieces of technology, a simple but important point can be established: at certain, crucial moments in history, private commercial interests, in the absence of government support, would not have become as intensely involved in certain long-term basic research and radical new concepts" (Flamm, 1988, p. 3). The teraflop objective has inspired parallel multi-microprocessor computers as the means for providing the next major jump in computer power.5
The teraflop objective has generated both attention and misunderstanding. Progress required building a number of machines large and fast enough to reward software researchers and application users with major gains in performance, thereby motivating them to develop the code that could make the high-performance machines useful. (See Appendix A for more information on the development of high-performance hardware and software and their interaction.) The costliness of this undertaking, compounded by the highly publicized financial difficulties of two entrepreneurial ventures, Thinking Machines Corporation (TMC) and Kendall Square Research (KSR), aimed at commercializing massively parallel computing systems, attracted criticism of the HPCCI.
However, that criticism appears largely misdirected. First, entrepreneurial ventures are always risky, and the two in question suffered from managerial weakness at least as much as questionable technology choices.6Contemporaneously, more established firms (e.g., Cray Research, IBM, Intel Supercomputing, Convex Computer, and Silicon Graphics Inc.; Parker-Smith, 1994a) have persevered, and others (e.g., Hitachi and NEC in Japan; Kahaner, 1994b, and Parker-Smith, 1994b) have entered or expanded their presence in the parallel systems market. Second, focusing attention on the high initial costs for stimulating development and use of parallel processing systems detracts from the achievement of successful proofs of concept and dissemination of new approaches to computation.
Although the teraflop objective was ambitious for the time scale set, it was intended as a driver and thus is best viewed as indicating a direction, not a destination; the need for progress in computing will continue beyond the teraflop capability.7In that respect, its appropriateness was affirmed by the 1993 Branscomb panel.8The teraflop objective has, in fact, served to focus attention on the task of combining and harnessing vast amounts of computer power from many smaller computers. The technology is now sufficiently developed that a teraflop machine could be realized today, although exactly when to do so should be left to the economics of users and their applications.9
Another direction-setting objective of the HPCCI was the achievement of data communications networks attaining speeds of at least 1 gigabit per second. Although by the mid1980s major telecommunications networks already had gigabit-plus trunk circuits in their backbones, the HPCCI was intended to lead to much broader deployment of and access to gigabit-speed networks connecting general-purpose computers. 10 This objective drove progress in switching, computing hardware and software, interfaces, and communication protocols.11 (See Appendix B.)
A third original objective related to applications of high-performance computing and communications technologies: to define and attack Grand Challenge problems. High-performance