Why Supercomputing Matters:
An Analysis of the Economic Impact of the Proposed Federal High Performance Computing Initiative

George Lindamood

George E. Lindamood is Vice President and Director of High-Performance Computing at Gartner Group, Inc., Stamford, Connecticut. He received his B.S., magna cum laude, in mathematics and physics from Wittenberg University and his M.A. in mathematics from the University of Maryland. He has more than 30 years of experience in the computer field, spanning academia, government, and industry, in activities ranging from research and development to international trade negotiations.

Introduction

On September 8, 1989, the Office of Science and Technology Policy (OSTP) published a report proposing a five-year, $1.9 billion federal High Performance Computing Initiative (HPCI). The goals of this program are to

• maintain and extend U.S. leadership in high-performance computing and encourage U.S. sources of production;

• encourage innovation in high-performance computing by increasing its diffusion and assimilation into the U.S. science and engineering communities; and

• support U.S. economic competitiveness and productivity through greater utilization of networked high-performance computing in analysis, design, and manufacturing.

In response to a Congressional request, OSTP and the Department of Energy, acting through Los Alamos National Laboratory, engaged Gartner Group, Inc., to develop a quantitative assessment of the likely economic impact of the proposed HPCI program over the coming decade. This study is proceeding in two phases.

In Phase I, which was completed in July 1990, two alternative scenarios (A and B), both depicting supercomputing through the year 2000, were developed. One scenario assumes full funding for the proposed HPCI program that would commence in FY 1992. The other scenario assumes "business as usual," that is, no additional federal funding above what is expected for HPCI-related activities now under way.

Phase II, which is the more important phase, was completed in September 1990. In Phase II, the two scenarios are extended to encompass the impact of the HPCI program, first upon selected industrial segments that are the major users of supercomputers and then upon the U.S. economy as a whole.

I will summarize the results of Phase I and describe the methodology employed in Phase II.

Phase I Methodology

During Phase I, two scenarios were developed. Scenario A assumes that current levels of HPCI funding will remain constant. Scenario B assumes full HPCI support, thereby changing the rate and direction of supercomputer development and utilization.

Scenario A

Our projection of the future of supercomputing is rooted in our understanding of the past, not only of supercomputing but also of other elements of the information industry. Over the last three years, we have developed a quantitative model that characterizes the information industry in terms of MIPS (millions of instructions per second), systems, and dollars for various classes of systems—mainframes, minicomputers, personal computers, etc., as well as the components of these systems, such as CPUs, peripherals, and software. Both the methodology and the results of this model have been applied in the development of the two alternative scenarios for the coming decade in supercomputing.

Basically, the model assumes that technology is the driver of demand because it is the principal determiner of both the overall performance and the price/performance of various types of information systems. Hence, future projections are based on anticipated technological advances,

interpreted through our understanding of the effects of similar advances in the past and the changing competitive conditions in the industry. Industry revenues are derived from these projections of price/performance, MIPS, and systems shipments, using average system price and average MIPS per system as "reasonability" checks. Historically, the model also reflects macroeconomic cycles that have affected overall demand, but there has been no attempt to incorporate macroeconomic forecasts into the projections.

Assumption 1. In modeling the supercomputer industry, we have assumed that supercomputer systems can be aggregated into three classes:

• U.S.-made vector supercomputers, such as those from Cray Research, Inc., Cray Computer Corporation, and (in the past) Control Data Corporation and Engineering Technology Associates Systems;

• Japanese-made vector supercomputers, such as those marketed by Nippon Electric Corporation, Hitachi, and Fujitsu; and

• parallel supercomputers, such as those made by Intel and Thinking Machines Corporation.

We assume that the price, performance, and price/performance characteristics of systems in each of these classes should be sufficiently uniform that we do not have to go into the details of individual vendors and models (although these are present in the "supporting layers" of our analysis). We do not anticipate any future European participation in vector supercomputers, but we assume no restrictions on the nationality of future vendors of the third class of systems.

Assumption 2. For our base scenario, we assume that the signs of maturity that have been observed in the market for vector supercomputers since 1988 will become even more evident in the 1990s after the current generation of Japanese supercomputers and the next generation of U.S. vector supercomputers—the C90 from Cray Research and the CRAY-3 (and -4?) from Cray Computer—have had their day.

Assumption 3. For parallel systems, however, we assume that the recent successes in certain applications will expand to other areas once the technical difficulties with programming and algorithms are overcome. When that happens, use of parallel systems will increase significantly, somewhat displacing vector systems—at least as the platform of choice for new applications—because of superior overall performance and price/performance. Growth rate percentages for millions of floating-point operations per second (MFLOPS) until the year 2000, installed, are shown in Table 1.

Assumption 4. We also assume that the price/performance of U.S.-made vector supercomputers will continue to improve, or decrease, at historical rates (about 15 per cent per year) and that the price/performance of Japanese-made vector systems will gradually moderate from 30+ per cent per year levels to 15 per cent per year by the year 2000. For parallel systems, we assume an accelerating improvement, to 20 per cent per year by the year 2000, in price/performance as a result of increasing R&D in this area.

Assumption 5. Despite the decrease in price per MFLOPS, average prices for supercomputer systems have actually increased a few percentage points per year historically. The reason, of course, is that average system size has grown significantly, especially because of expanded use of multiprocessing. We assume that these trends will continue for vector systems, albeit at a slowed rate of increase after 1995, because of anticipated difficulties in scaling up these systems to ever-higher levels of parallelism. For parallel systems, technological advances should lead to accelerated growth rates in processing power, resulting in systems capable of one-TFLOPS sustained performance by the year 2000. Growth rate percentages for average MFLOPS per system are shown in Table 2.

Assumption 6. Finally, we assume that retirement rates for all classes of supercomputer systems will follow historical patterns exhibited by U.S.-made vector systems.

These assumptions are sufficient to generate a projection of supercomputer demand for the next 10 years:

• The number of installed systems will more than triple by the year 2000. Table 3 shows how this installed base will be divided, as compared with today. The number of supercomputers installed in Japan will exceed the number installed in the U.S. after 1996.

• Installed supercomputer power (measured in peak MFLOPS) will increase more than 125-fold over the next decade, from almost 1.4 million MFLOPS in 1990 to over 175 million MFLOPS in 2000. (However, this is substantially less than the growth rate in the 1980s—from about 4000 MFLOPS in 1980 to 340 times that amount in 1990.) Of the MFLOPS installed in 2000, 90 per cent will be parallel supercomputers, two per cent will be U.S.-made systems, and eight per cent will be Japanese-made systems.

• As shown in Table 4, the "average" vector supercomputer will increase about 10 times in processing power, whereas the "average" parallel system will increase about 60 times over the decade. Average supercomputer price/performance will improve by a factor of 25,

mostly as a result of increased usage of parallel systems that have more than 10 times better price/performance than vector systems.

• Annual revenues for vector supercomputers will peak at just under $3 billion in 1998. Revenues for parallel systems will continue to grow, surpassing those for vector systems by 1999 and exceeding $3.1 billion by 2000.

Scenario B

For this scenario, we assume that the federal HPCI program will change the direction of high-performance computing (HPC) development and utilization and the rate of HPC development and utilization.

Assumption 1. As in Scenario A, supercomputers are grouped into three classes in Scenario B:

• U.S.-made vector supercomputers,

• Japanese-made vector supercomputers, and

• parallel supercomputers.

Assumptions 2 and 3. We assume that demand for supercomputer systems of both the vector and parallel varieties will be increased by the HPCI program components concerned with the evaluation of early systems and high-performance computing research centers. All funding for early evaluation ($137 million over five years) will go toward the purchase of parallel supercomputers, whereas funding for research centers ($201 million over five years) will be used for U.S.-made vector and parallel supercomputers, tending more to the latter over time. We also assume that federal funding in these areas will precipitate increased state government expenditures, as well, although at lower levels. Although all of these systems would be installed in academic and government facilities (primarily the former), we also postulate in Scenario B that the

technology transfer components of HPCI would succeed in stimulating industrial demand for supercomputer systems. Here, the emphasis will be more on U.S.-made vector systems in the near term, although parallel systems will also gain popularity in the industrial sector in the late 1990s as a result of academic and government laboratory developmental efforts supported by HPCI.

Assumption 4. This increased demand and intensified development will also affect the price/performance of supercomputer systems. For U.S.-made vector systems, we conservatively assume that price/performance will improve one percentage point faster than the rates used in Scenario A. For parallel supercomputers, we assume that price/performance improvement will gradually approach levels typical of microprocessor chips and RISC technology (that is, 30+ per cent per year) by the year 2000.

Assumption 5. The increased R&D stimulated by HPCI should also result in significantly more powerful parallel supercomputers, namely, a TFLOPS system by about 1996. However, we do not assume any change in processing power for vector supercomputers, as compared with Scenario A, because we expect that HPCI will have little effect on hardware development for such systems. (This is distinct, however, from R&D into the use of and algorithms for vector systems, which definitely will be addressed by HPCI.)

Assumption 6. We assume that retirement rates for supercomputer systems of all types will be the same as in Scenario A.

As before, these assumptions are sufficient to generate a projection of supercomputer demand for the next 10 years:

• The number of installed supercomputers will approach 2200 systems by the year 2000. Table 5 shows how this installed base will be divided, as compared with Scenario A.

• Particularly noteworthy is the difference between these two scenarios in terms of U.S. standing relative to Japan. In Scenario A, Japan takes the lead in installed supercomputers, but in Scenario B, the U.S. retains the lead.

• Installed supercomputer power (measured in peak MFLOPS) will be increased by a factor of more then 300, to over 440 million MFLOPS, by the year 2000 (which is slightly less than the rate of growth in the 1980s). Of the MFLOPS installed in 2000, 96 per cent will be parallel supercomputers, one per cent will be U.S.-made vector supercomputers, and three per cent will be Japanese-made vector supercomputers.

• As shown in Table 6, the "average" vector supercomputer will increase about 10 times in processing power, whereas the "average" parallel system will increase nearly 125-fold over the decade. Average supercomputer price/performance will improve by a factor of 55.

• Annual revenues for vector supercomputers will peak at just over $3 billion in 1998. Revenues for parallel systems will continue to grow, surpassing those for vector systems by 1997 and exceeding $5 billion in 2000.

The differences between Scenarios A and B, as seen by the supercomputer industry, are as follows:

• 17 per cent more systems installed;

• almost three times as many peak MFLOPS shipped and two and one-half times as many MFLOPS installed in 2000;

• 39 per cent greater revenues in the year 2000—an $8 billion industry (Scenario B) as opposed to a $5 billion industry (Scenario A); and

• $10.4 billion more supercomputer revenues for the 1990–2000 decade.

In addition to these differences for supercomputers, HPCI would cause commensurate increases in revenues and usage for minisupercomputers, high-performance workstations, networks,

software, systems integration and management, etc. However, the largest payoff is expected to come from enhanced applications of high-performance computing.

Phase II Methodology

To estimate the overall economic benefit of HPCI, we have sought the counsel of major supercomputer users in five industrial sectors representing a variety of experience and sophistication:

• aerospace,

• chemicals,

• electronics,

• oil and gas exploration and production, and

• pharmaceuticals.

Our assumption is that supercomputers find their primary usage in R&D, as an adjunct to and partial replacement for laboratory or field experimentation and testing—for example, simulating the collision of a vehicle into a barrier instead of actually crashing thousands of new cars into brick walls. Hence, high-performance computing enables companies to bring more and better new products to market and bring new products to market faster.

In other words, high-performance computing improves R&D productivity. Even if there is no other benefit, the use of HPC, which in turn affects overall company productivity in direct proportion to the share of expenditures for R&D, provides a way to determine a conservative estimate of productivity improvement, as shown by the following steps:

• Scenarios A and B are presented to company R&D managers, who are then asked to give estimates, based on their expertise and experience, of the change in R&D productivity over the coming decade.

• For both scenarios, these estimates are translated into overall productivity projections, using information from the company's annual report, for the ratio of R&D spending to total spending.

• Productivity projections for several companies in the same industrial sector are combined, with weightings based on relative revenues, to obtain overall Scenario A and B projections for the five industrial sectors identified above. At this point, projections for other industrial sectors may be made on the basis of whatever insights and confidence have been gained in this process.

These productivity projections are interesting in and of themselves, but we do not intend to stop there. Rather, we plan to use them to drive an input/output econometric model that will then predict the 10-year change in gross national product (GNP) under Scenarios A and B. By subtracting the GNP prediction for Scenario A from that for Scenario B, we expect to obtain a single number, or a range, that represents the potential 10-year payoff from investing $1.9 billion of the taxpayers' money in the federal HPCI program.