High Performance Computing Export Controls: Navigating Choppy Waters

By Peter Wolcott, Seymour Goodman, Patrick Homer

Communications of the ACM, Nov. 1998, vol. 41, no. 11, pp. 27-30.

ACM COPYRIGHT NOTICE. Copyright © 1998 by the Association for Computing Machinery, Inc. Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from Publications Dept, ACM Inc., fax +1 (212) 869-0481, or permissions@acm.org.

The debate over export controls of high-performance computing (HPC) never goes away. In October, 1995 the President announced the second major revision to the HPC export control policy of his administration. Since then, press reports and opinion pieces [1-6], a Government Accounting Office report [7], and isolated illegal sales and diversions of HPC hardware have repeatedly brought this issue back into public debate. The past year and a half have seen Congressional hearings and new legislation tightening licensing requirements. This debate, however, has masked two important aspects of the export controls since 1995. First, an attempt has been made to put the policy formation process on a more rational, explicit, defensible, and sustainable foundation. Second, much of the debate has centered on applications that can be performed successfully using relatively low and readily available levels of computing by today’s standards, most notably nuclear weapons development. Such applications can no longer justify an effective export control regime; the debate must find others that can.

A firmer foundation

In the nearly half century since HPC export controls were first formulated, there has been broad agreement about the general policy objectives, and bitter disagreement over implementation details. The main objective is to slow development of certain applications by particular nations by limiting their access to the computing hardware needed. The policy has been implemented by subjecting the sale of computers whose performance exceeds a specified threshold to extraordinary licensing conditions. The question of which computers should be subject to such licensing has been hotly debated. Because the technologies, markets, and international geopolitical landscape change continuously, the policy must be revised periodically.

The policy formulation process has recently made significant progress towards becoming more rational, explicit, defensible, and sustainable than has been the case throughout its history. Four major developments have made this possible.

First, there exists a unifying framework reflecting the salient technical issues and viewpoints that can guide policy-makers toward selection of control thresholds. This framework, developed in [8,9] and elaborated in [10], posits and tests three basic premises:

1. There are problems of great national security importance that require HPC for their solution, and these problems cannot be solved, or can only be solved in severely degraded forms, without such computing assets.
2. There are countries of national security concern who have both the scientific and military wherewithal to pursue these or similar applications.
3. There are features of the necessary computers that permit effective forms of control.

Arguably, all three premises held during most of the Cold War. With changes in threats to U.S. national security interests and in HPC technologies, markets, and applications, the premises no longer hold at Cold War levels. The new framework helps establish a range of threshold choices for which the basic premises do hold at a given point in time. It can accommodate any data that is relevant to the discussion, e.g. technical details about computing systems, market data, computational requirements of national security applications, uses of HPC in foreign countries, or foreign HPC vendors. Consequently, one may demand that the loudest voices in the discussion "put up their data."

Critical to the policy is the ability to satisfactorily monitor and regulate the diffusion and use of HPC hardware throughout the world. While the precise determination of what kinds of technology are controllable is subject to considerable debate, controllability is not solely a function of foreign availability in the traditional sense. Historically, "foreign availability" referred to foreign production sources, not subject to export control regulations of the United States and its allies. U.S.-based companies overwhelmingly dominate the international HPC market. Yet even their products may be considered uncontrollable if they are sufficiently numerous, inexpensive, small, easily installed and upgraded, or available on secondary markets. Even the most conservative voices must agree that the Intel-compatible microprocessors –manufactured in tens of millions of units per year only by American companies (Intel, AMD, Cyrix) or their foreign licensees – are uncontrollable. However, more powerful systems such as workstations and some multiprocessors are also likely to be, in practice, uncontrollable.

Second, HPC vendors now report to the Department of Commerce all sales of systems above 2,000 Mtops. Thus, it is possible to estimate more objectively the cost to U.S. industry of various proposed control thresholds.

Third, the regulations now divide the countries of the world into four "tiers," each with different licensing requirements (Table 1). By far the largest foreign HPC markets are in those countries of little threat to U.S. national security. In 1996 and 1997, the number of systems between 2,000 and 7,000 Mtops exported to Tier 1 countries was over 10 times greater than the number exported to Tier 3 destinations. Above 7,000 Mtops, licensing restrictions limited the number of Tier 3 exports to just 0.5% of the number shipped to Tier 1 countries. By permitting exports to Tier 1 countries to take place under general license, policy makers have done little to compromise U.S. national security interests.

Table 1: Tiers and associated HPC export licensing limits

Fourth, policy makers have committed to revisiting the policy regularly. Previously, the policy was revised only when the pressure for change became intolerable. Consequently, stakeholders were forced to speculate on trends far into the future and to take strident positions. The time between revisions is still longer than what it should be, but when industry stakeholders are confident that regular reviews will take place, they may be less inclined to take extreme positions.

Taken together, these four developments have caused the level of agreement among the major stakeholders to expand considerably. This point is not easily seen in the discussion in the press, which appears driven by the most strident views.

National security applications

There must be applications satisfying the first basic premise to justify the policy. Nuclear weapons design has been frequently held up as such an application. The nuclear weapons in the current U.S. (and Russian) stockpiles were designed using nothing more powerful than a single-processor Cray X-MP or Y-MP. During the 1980s, these systems were the most powerful in the world and had features that made them readily controllable.

Now, however, Cray Y-MP levels of computing power are widely available in arguably uncontrollable forms. For example, the major RISC vendors now annually manufacture hundreds of thousands of microprocessors each more powerful than a single-processor Cray Y-MP (500 Mtops), and Intel is reaching this level this year. Among scientists and other officials at the Department of Energy and its weapons labs, there is widespread agreement that when test data are available, computing above this level is no longer on the critical path in designing nuclear weapons. When test data are not available to validate codes, more powerful computers are of limited benefit. Certainly, given quality test data, marginally more computing power is marginally more beneficial, but systems of several thousand Mtops offer no truly exceptional and discontinuous advantages for nuclear weapons design over a few easily networked lesser machines.

If nuclear weapons design no longer can justify the policy, what applications do? Applications that fulfill both parts of the first premise fall into two broad categories: operational applications, and research and development applications. The former include weather forecasting for military operations, and very sophisticated cryptography and signal processing. R&D examples include large-scale forces modeling, effects of weapons on structures, radar cross-section calculations, and shallow water submarine maneuvers.

Such applications utilize machines that are controllable. One can also argue that they are important for national security. However, it has been difficult to find constituencies willing to fight for these applications in the export control debate. In conducting our research, we were surprised at how difficult it was to find individuals within the Departments of Defense or Energy who were willing and able to argue for a continuation of export controls at the workstation or modest multiprocessor levels to protect applications within their spheres of activity. Without such constituencies, the first basic premise is weakened.

Unresolved questions

A number of issues remain unresolved. Parallel architectures based on commercial microprocessors dominate today’s multiprocessor markets. These systems are increasingly easily scalable across a broad performance range. It is possible for an end-user to acquire legitimately a lesser-powered computer, then add CPU and memory boards to increase the system’s performance above a control threshold. To guard against this, the current policy relies on a combination of U.S. government and vendor efforts to monitor systems after shipment to ensure that they are in compliance. This simplifies the licensing process, but reliable post-shipment verification is difficult. An alternative approach is to consider the degree to which system performance can be scaled as part the license application review. Systems that are easily scalable without vendor support might be treated more strictly, especially if they are to go to Tier 3 countries. The fundamental decision facing policy makers is whether it is better to control for scalability before, or after system installation.

A second major issue is what, ultimately, is the goal of the policy? There are generously perhaps a few hundred Tier 3 and 4 facilities capable of using HPC systems effectively for applications of national security concern. Is the objective to keep these often well-heeled facilities from acquiring even one or two computing systems? Or, is the objective to minimize the total amount of computing capability entering the military and civilian economies of these countries? For the first option, efforts to control computing systems with installed bases in the tens of thousands—deskside multiprocessors, for example—are likely to fail; individual systems from this large base are easily acquired on the world market in small numbers by such organizations. In its final years, the Soviet Union successfully imported and used large numbers of 286-based PCs and U.S-made VAX-class minicomputers, which at that time were still subject to export control and enforcement through the 16-nation Coordinating Committee for Multilateral Export Controls (COCOM). For the second option, controls on such systems may effectively slow the aggregate flow of technology into a country. But in either case, the most worrisome facilities will be able to acquire computers whose current controllability is less than that of VAXes in the late 1980s, without much difficulty or delay. Furthermore, the demise of COCOM essentially leaves the United States alone as a serious controller of computer exports.

A third unresolved issue is how foreign facilities of concern might employ HPC technologies to the detriment of U.S. national security interests. For decades, a "mirror-imaging" analysis has been employed in which one assumes that foreign entities are likely to pursue applications in a manner comparable to U.S. practitioners. Mirror-imaging analysis is not completely accurate, but the real questions are whether it is good enough, and whether there is an alternative? An asymmetry between U.S. and foreign pursuits is significant only if a) a foreign country pursues applications not pursued by the U.S., or b) if foreign entities are able to use lower levels of HPC technology far more effectively than their U.S. counterparts. Given the great depth and breadth of experience and support of the U.S. national security HPC community, the first point is unlikely. The second point may be true in isolated instances; however it is extremely difficult and expensive to acquire such information. Mirror-image analysis becomes the only viable general alternative. However, the current analysis framework can accommodate such information as it becomes available.

For good reasons, there is great concern over the proliferation of nuclear, chemical, and biological weapons of mass destruction (WMD). Perhaps as a result of Cold War history there continues to be a misperception that HPC is a major enabler of this threat. This yields an exaggerated view both of the importance of computing for WMD proliferation, and of our ability to deny widely available levels to undesirable end users. Our energies should be focused on other, more effective measures.

[1] Lelyveld, M. S., "White House backing off veto over supercomputers; Senate set to vote on tighter export controls," Journal of Commerce, Nov 4, 1997, p. 3A.

[2] Blustein, P., "Computer Evolution: Faster Than a Speeding Export Curb," Washington Post, Jul 3, 1997, p. E01.

[3] Reinsch, W. A., "US Export Controls--Adequate and Improving," Journal of Commerce, Aug 14, 1997, p. 8A.

[4] Milhollin, G., "Should We Sell Supercomputers to Algeria," New York Times, Apr 24, 1998, p. A27.

[5] Clark, B., "Computer exports ‘aiding China nuclear weapons’," Financial Times, Jan 13, 1998, p. 04.

[6] Hedges, S. J., "Trade Over-ruled U.S. Security, Analyst Asserts," Chicago Tribune, Jun 26, 1998, p. 11.

[7] Hedges, S. J., "GAO Audit Criticizes Computer Sales to China; Commerce Official Defends Easing Rules," Chicago Tribune, Jul 30, 1998, p. 1.

[8] Goodman, S. E., P. Wolcott, and G. Burkhart, Building on the Basics: An Examination of High-Performance Computing Export Control Policy in the 1990s, Center for International Security and Arms Control, Stanford University, Stanford, CA, 1995.

[9] Goodman, S., P. Wolcott, and G. Burkhart, An Examination of High-Performance Computing Export Control Policy in the 1990s, IEEE Computer Society Press, Los Alamitos, 1996.

[10] Goodman, S. E., P. Wolcott, and P. Homer, High-Performance Computing, National Security Applications, and Export Control Policy at the Close of the 20th Century, Stanford University, Stanford, CA, 1998. http://www.stanford.edu/group/CISAC/test/pub/recent.html