Dear Editor:
I concur with Fran Berman's conclusions in her recent feature article in HPCWire [440274]. Indeed, we must set strategic priorities and concrete goals to ensure US technical leadership. However, I suggest that we must understand and acknowledge, not rationalize, the decline of the US HPC industry over the past 15 years and embark upon a much more specific new National Economic Security Initiative for the improvement of the US HPC industry. This, in my opinion, is mandatory if we hope to continue to lead the world in the development of intellectual property, the most basic commodity of our future international competitiveness.
What we have is a market where everyone, even IBM in my opinion, is losing money at the high end. There is just not enough high-end revenue to go around to sustain profitable operations, much less fund product development. Fortunately, IBM and HP have sufficient sales at the low-end to carry their losses from the high-end. Sun has virtually exited the high-end. The same may be said shortly for HP, given its new CEO's reputation for driving profits and the recent announcement of up to 14,500 personnel cuts. It's interesting to note that in 1993, with a 70 percent high-end market share, Cray sold C-90s for $2,500 per megaflop and, of course, made money. Today, the going rate for cluster systems is as low 25 cents a megaflop with no single dominant supplier. Rather than develop real technical differentiation, most vendors are seeking market visibility by trying to impress the market with “macho FLOPS” on the top 100 list.
What has brought us to this state? First, most real world, “bread and butter” applications that used to require supercomputers can be run on simple midrange clusters. Most production applications have not been expanded/scaled up as systems have increased in size. The systems might not be efficient, but they are good enough. And that best sums up the situation. We are now in a “Good Enough” environment. Even the pursuit of new science seem to be changing Whereas new science usually used to require supercomputers, today new science is often pursued by running many capacity scale jobs to develop a panorama of results from which innovation is derived.
Second, the US government, that has led the high end market for so long, has failed, despite the HPCS program, to provide the traditional leadership to keep the US high end computer industry vibrant. At a recent Council on Competitiveness HPC Users Meeting, DreamWorks Animation produced a short film to emphasize the importance of HPC to the national competitiveness interest. I hope it gets wide dissemination among decision makers in the Administration and on Capitol Hill. Unfortunately, as articulate as it is, it fails to point out two points: first, the US high-end supercomputer industry is in very poor financial health; and second, our prime international computer competitor, Japan, and I believe China and India as well, consider HPC as their prime technical priority for international scientific competitiveness.
Despite the fact that we live in a world of conflicting priorities, Congress must look past its usual vision of the long term, the next election, and recognize that the long-term future of this country is embodied in the development of intellectual property, not in manufacturing, even high tech manufacturing. Congress must make a commitment for this country to continue to lead the world in intellectual property through the investment in the most fundamental enabling technology for the development of IP, high performance computing. Our international competitors are!
So, to augment Fran Berman's more general call for national scientific goals and objectives, here are my specific recommendations for a new National Economic Security Initiative for the future:
1. Congress must recognize the importance of balance between hardware and software as we pursue PetaFLOP computing by funding over the next five years development of application software, middleware, and tools to allow users and independent software vendors to scale existing programs or create new ones to take advantage of PetaFLOP systems when available. To accomplish this, $300 million to $500 million per year for five years will be necessary. Because we cannot afford to wait, the funds for FY 06 should come from the Office of Science at DOE by reprogramming funds within existing programs. For the following years, new funds should be authorized and appropriated for DOE.
2. Because successful PetaFLOP software development will require a broad effort among industry, academia, hardware and software vendors, and government, Congress must fund a major new HPC infrastructure initiative throughout academia, the supercomputer centers, and the national labs, to provide a broad network of unclassified systems for this software development. As we approach petaFLOP computing, the US scientific community needs a large network of fifty to one hundred large new systems on the order of 100 TeraFLOPS or more to provide much broader access to real high end HPC. Installations should begin in FY06 with the objective that all be completed by mid FY 08, at which time refresh of the older systems should begin. My estimate of the cost of this initiative is one billion dollars per year for five years and must be new money to augment existing programs, especially DARPA's HPCS program.
3. Because of the concern regarding lack of students' interest in math and science, new initiatives with special student incentives from the National Science Foundation and/or the Department of Education are necessary to ensure the long term supply of talent to fund the US intellectual property machine of the future. I will reserve judgment on the scope of this recommendation but encourage Congress to pursue this issue as an investment in America's future.
4. Government should recognize that the US HPC industry must be economically competitive around the world. Until a new disruptive technology emerges, commodity processor based clusters will, because of their economic advantage, dominate the market and should be the point of government financial emphasis. If national security demands dictate continuance of custom architectures for specific applications (the submarine model), so be it; however these demands should be treated separately from mainstream technical commercial computing, and not included in the initiatives suggested above as they are not now economically competitive. Finally, for these suggestions to become reality, they must rise in the government funding priority stack. American business and academia, as well as the relevant senior executives in government, must make these issues known to Congress for their action.
Sincerely,
Charles. W. Hayes
Readers can send further comments to [email protected].