Scientists Must Speed Up Transfer Of Supercomputing Technology

Although the United States continues to be recognized as a world leader in science, our industries have increasingly been buffeted by global competition. Other nations, particularly Japan, have often been swifter than we in moving their own scientific and technological advances from the laboratory to the marketplace. Supercomputing offers an opportunity for the United States to address these problems, using a homegrown technology in which we still maintain a lead. For example, circuit simulation that might take two months on a minicomputer can be done in about a day on a supercomputer. Such quick turnarounds can make the difference between whether work is done computationally or done through costly and lengthy prototyping and product testing. Further improvements in cost and performance could make these powerful tools available to a wide range of industrial users, including even small and mid-size companies. Supercomputers available by the mid-1990s are expected to be at least 100 times faster than mid-1980s machines. And price/performance gains will also result from design or architecture innovations and improved software. As supercomputer capabilities increase, they will make possible new applications, raising prospects for fundamental changes in product development as well as many other aspects of manufacturing. Already there have been major shifts in product design, manufacturing processes, and office work based on the use of much lower performance computers and workstations. But high performance computing has so far remained exclusively the province of an elite cadre of scientists at major universities and federal research laboratories and within a few of our largest firms, where large and complex research initiatives are being pursued. Some efforts have recently been made to broaden the supercomputer user base through shared access, allowing even small users to participate. But these efforts fall far short of what will be needed if high-performance computing is to become an integral part of the American industrial infrastructure. There are three basic barriers to wider use of supercomputers by industry: the scarcity of appropriate skills, the lack of appropriate software, and the perceived high cost of supercomputer access. Supercomputers are sophisticated tools, and making the most of them requires expertise; this is particularly critical for high-performance computing because there is a shortage of the powerful application tools that have accumulated for conventional computers. Expertise in the handling of numerically intensive problems is also essential. Because supercomputers process large volumes of data and generate large volumes of output, users must be skilled in the design of algorithms and complex models. Unfortunately, expertise relevant to supercomputing is a scarce commodity. Firms with limited resources therefore have a tough time competing against well-funded corporations and research institutions in attracting the relatively small number of experienced supercomputer users. Further compounding the problem is the lengthy period needed for both formal education and on-the-job training. Even if efforts to provide the needed skills were stepped up, it would take many years to enlarge the pool of adept users. Unlike software for personal computer users, who can go to a local store or a mail-order catalog and select from thousands of software packages, application software for supercomputers is much harder to find. Many applications are user-specific, and there are only a few hundred commercially available packages. Supercomputer developers are strong contributors to this problem. Vendors focus on achieving ever higher levels of performance (for example, more and more millions of instructions or operations per second) for high-level scientific and engineering users. And they generally provide systems software that is as sophisticated as the hardware. But broadening access will require greater consideration of how improved performance can best be channeled into productive applications by a wide range of users; software must be more use friendly. Supercomputers are expensive-fully configured systems still cost several million dollars-which discourages many potential users. Specialized facilities, maintenance (often from onsite vendor personnel), operator skills, and end-user experience are all required to take advantage of these complex systems. Thus purchases of supercomputers have been limited to relatively large or affluent organizations; in the United States, which is far ahead of any other nation in the use of supercomputers, fewer than 300 are installed. Even large firms balk at the necessary investment, just as many of them have resisted factory automation and other large-scale implementations of computer technology. And many companies that do see its potential may continually put off the investment in anticipation of better systems in the future. But there is an alternative. "Shared access" is growing as a means of spreading costs, much as time-sharing has served mainframe computing. Supercomputer centers provide remote as well as local access, with remote subscribers tapping the machines via private or networked data links. Supercomputer centers also offer skilled staff assistance, including applications consulting, software development, and training. Unfortunately, relatively few centers exist today, and their capacity is already strained in support of a largely academic body of research users. All these barriers are symptoms of a more fundamental problem, a flawed process of technology transfer to the marketplace from basic research environments in government and academia. There has been some success in the defense arena, where federal agencies have provided incentives for academia and industry to collaborate and share technology. here, the transfer process is catch-as-catch-can. Despite the unsystematic nature of technology transfer in this country, successful innovations have still been made. In supercomputing, however, the impediments are so great that the United States could well find itself lagging behind countries with better-orchestrated programs. England, France, and West Germany, as well as Japan, have already stepped up funding for supercomputer research and applications development. Such initiatives helped motivate the National Science Foundation to establish a Supercomputer Research Center program. NSF's program supports five facilities-the Pittsburgh Supercomputer Center, the Cornell National Supercomputer Center, the John von Neumann National Supercomputer Center in Princeton, N.J., the National Center for Supercomputer Applications at the University of Illinois, and the San Diego Supercomputer Center. All of these facilities provide access to high- performance computers for researchers in science and engineering; they also support development of advanced research computing tools and techniques. The program is a great step in support of academic research, but it falls short of what is needed to invigorate technology transfer from universities and other basic research settings- where so much of the supercomputing know-how is found-into industry. One critical limitation in nurturing industrial applications has been that no more than 10 percent of NSF center resources could be devoted to proprietary commercial ventures. Although this restriction has been somewhat relaxed, it has still effectively meant a strong focus on basic science, which is of limited industrial interest beyond top-tier firms. In any case, the affiliates programs offered by the centers require substantial monetary commitments (ranging up to a few million dollars) in return for large blocks of time, which further discourage interest by smaller firms. Thus, although center personnel report that some 6,000 people have accounts on supercomputers at these centers, they also acknowledge that few of them are regular or sophisticated users, and many are dedicated to basic scientific research. In order to make supercomputers more widely available- especially to smaller firms-we need collaborations to bring industry, academia, and government together. One such partnership is the "incubation initiative," drawing on the existing framework of state economic development and education programs with access to local universities for training and consulting. Some states already have begun to develop facilities, typically affiliated with state universities, that could provide a basis for expanded industrial application of supercomputers. But because these programs have primarily supported academic researchers, additional steps may be needed to protect the proprietary interests of industrial clients. A program that would be of great value to industry-especially to small to medium-sized firms that could not afford the equipment or skilled personnel used by larger companies-would involve a state supercomputer center offering applications across the whole manufacturing cycle. Such programs do not currently exist. But it is instructive to suggest how entire regions could benefit from them. Michigan, for example, could provide sophisticated support to its motor vehicle and heavy equipment industries. Everything from simulation of a prototype auto body to parts design to production modeling could be available. Supercomputer analysis for the Ford Taurus, for example, eliminated the need for multiple prototypes, saved vast sums of money through collision-damage modeling, and helped to increase fuel efficiency. Modeling might be done for individual parts or processes, or even combined so that suppliers and assemblers might jointly explore tradeoffs. This would build on what has already begun: Automakers are requiring suppliers to accept and exchange parts design information in computer-aided design form and to transmit invoices and shipping notices electronically. Such a program might evolve from a pilot project being considered by the staff of the New York Center for Theory and Simulation in Science and Engineering, which also houses the Cornell National Supercomputer Center. The project would serve small and medium-sized companies by training corporate scientists in the application of advanced computing techniques to product development. Corporate participants would each pay $2,000, a fee low enough to open access to a wide range of firms. Through a preliminary survey, Cornell identified such prospective clients as a 75-employee chemical company interested in supercomputing for reactive polymer work and the development of new, heat resistant materials; an environmental laboratory seeking supercomputer access to develop pollutant transport models and data bases; and a six-person precision instrument firm in need of sophisticated design and modeling capabilities. The federal government's best role in all this appears to be as a source of seed funds and assistance with capital acquisition. But federal help should be tied to program plans, involving regional industry and local resources, that promise self-sufficiency after a reasonable start-up period. The formula might vary from state to state depending on needs, but programs should operate within the context of existing state or regional economic development activities. It is important to consider where the federal support should come from. The existing NSF centers are nearing saturation just in meeting their research objectives. To distinguish any new efforts from the present NSF program, and to underscore the commercial intent, the appropriate agency might be the Department of Commerce, possibly in the Economic Development Administration (although the EDA is a perennial target of cost-cutters), or the National Institute of Standards and Technology. Coordination with NSF and its centers, and with other concerned agencies, could help propagate information about new programs to potential participants while also building support within the government. In any case, it is hoped that growing concern in Washington about the competitiveness of U.S. industry will help rally the political commitment needed to ensure adequate funding. Sometimes we in the United States are more obsessed with going for the home run than with the fundamentals that win games. Our focus is on who gets the most prestigious awards; whose grand challenges are the grandest. Although all of this will have significant payoff, there are other questions that are less glamorous but possibly more important: How do we design higher-quality cars? How can we bring chemicals to market faster? And how do we make our manufacturers more productive? This country's experience as the pioneer in developing supercomputers, and as an innovator in applying them, can provide a potent competitive advantage. The challenge is to use that advantage while we still have it. As foreign competition grows, aided in part by aggressive industrial support programs overseas, it is clear that we need leadership and vision to ensure that U.S. firms make the most of new technology to help them compete. We may never approach the single-minded purpose of some other countries, but through small and large collaborations, the result may be greater. Given the individualistic bent of U.S. enterprise, such collaboration needs encouragement to get started. The NSF supercomputer centers have shown how this can be done for basic research; additional and more appropriate centers are needed to reach the majority of firms and their product and process needs. Irving Wladawsky-Berger is vice president of the Data Systems Division and general manager of the Kingston complex of the IBM Corporation. Marjory S. Blumenthal is executive director of the Computer Science and Technology Board of the National Research Council's Commission on Physical Sciences, Mathematics, and Resources. This essay first appeared in Issues in Science and Technology (Vol. 5, No. 3, 1989), published by the National Academy of Sciences.

Scientists Must Speed Up Transfer Of Supercomputing Technology

Interested in reading more?

Become a Member of