In this issue, I highlighted a recent commentary written by Prof. Hans Werner Meuer, who writes about the 15 year anniversary of the TOP500 project. Meuer did a great job at recounting the history of the list and provided some interesting anecdotes of his favorite machines. Although in the past, I’ve offered criticism of the list, in the feature article I focused on some of the more practical aspects of the TOP500 project that Meuer discusses. That includes using the list as a historical record for tracking the evolution of high performance computing over the last decade and a half, and to project its future.
One of the more positive effects of the TOP500 project — and one that Meuer doesn’t mention — is that it helps to popularize the notion of supercomputing to the wider industry and maybe more importantly, to the general public. While this might seem superficial, the twice yearly list at least provides valuable PR for the almost invisible HPC community.
Save for the occasional article in the mainstream media about how supercomputers have predicted climate changes or discovered some mystery of the universe, most of HPC is hidden from public view. The missing element in most stories about supercomputers is how they relate to the human condition at the scale of the individual. It’s not that climate modeling, derivative pricing or seismic simulations are not worthwhile applications. But linking supercomputing to personal applications would inspire a new generation of scientists and engineers.
Consider this. In Meuer’s 15-year TOP500 retrospective mentioned above, he talks about some of his favorite supercomputers, one of which was “Deep Blue,” the IBM machine that bested world chess champion Garry Kasparov in 1997. Arguably the most famous supercomputer ever built, Deep Blue ranked a modest 259 on the 9th TOP500 list.
“This system, named Deep Blue, was installed at the IBM Watson Research Center in Yorktown Heights and had a best Linpack performance of 11.37 Gigaflop/s; it was an IBM SP2 P2SC with 32 processors and a clock rate of 120 MHz. But the floating point performance was not what really mattered: Each of the 32 processors was equipped with 15 special-purpose VLSI chess chips. Deep Blue was the first chess computer to beat a reigning world chess champion, Garry Kasparov. Ten years after this event, no chess player stands a chance against any kind of computer, not even against a simple home computer. One year ago, in November/December 2006, Deep Fritz played a six-game match against reigning world chess champion Wladimir Kramnik in Bonn. Deep Fritz won 4-2.”
The chess matchup inspired a book by Deep Blue’s designer Feng-hsiung Hsu, “Behind Deep Blue,” and a 2003 documentary, “Game Over: Kasparov and the Machine.” Since the machine was retired after beating Kasparov, no supercomputer has quite captured the public imagination the way Deep Blue did. Why? Almost everyone can relate to playing a game of chess, whether or not they have actually done so.
Contrast this with Meuer’s number one favorite, Intel’s ASCI Red, the first teraflop machine. Although, one of the most legendary supercomputers ever built, it’s almost unknown outside of the HPC community. This machine was used to provide simulation capabilities for nuclear weapons — an eminently useful task, but one not likely to be taken up by the average citizen (one hopes).
Software with elements of artificial intelligence (AI), like chess playing, constitute some of the most compelling computing applications to the public. And while AI and supercomputing might seem like a natural fit, for a variety of reasons, the two communities never really hooked up. Today, the AI field remains fragmented with a lot of competing paradigms and not much to show for it. Most of the work never made it out of the lab or the lecture hall. After 50-plus years of research, a lot computer techies sneer at the whole AI meme.
Yet it is these types of applications that have the potential to excite a new generation of technologists and motivate them to become involved in something more transformative than say quantum chromodynamics (not that there’s anything wrong with QCD groupies). The commercialization of AI could help place supercomputing back into the public eye. We’re not necessarily talking about “big iron” here. A lot of machine intelligence, though, will likely require massive levels of parallelism, in a tightly-coupled architecture (think manycore). Intel is currently developing a software model for terascale computing platforms that would enable such applications. Called RMS — for recognition, mining, synthesis — Intel believes this technology could be the basis for the killer apps of the next decade.
In Ray Kurzweil’s 1990 book, “The Age of Intelligent Machines,” he talked about a number of AI systems that would be developed in the 21st century. At a time when Pac-Man was the cutting edge in computer games, he described computer-generated animation twenty years into the future with uncanny accuracy: “Reasonably lifelike video images of human faces … completely synthesized and animated.” He also predicted the triumph of a computer over the top human chess player by 1998 — actually underestimating Deep Blue’s accomplishment by one year.
In his book, Kurzweil mentioned a number of other applications that have not yet come to pass, including the cybernetic chauffeur, the intelligent assistant, the intelligent answering machine, and the translating telephone. The latter is a machine that would automatically translate the language of the two parties as they spoke in real-time. The technology required to do so includes automatic speech recognition, language translation, and speech synthesis. All three existed at the time the book was written 18 years ago, but not nearly at the level of sophistication they are today. With the continued improvement in microprocessor power and software, it is certainly conceivable that such technologies will be incorporated into virtually every phone and phone-like device within the next 10 years.
The effects of ubiquitous language translation would be enormous. That single application would not only greatly accelerate economic and social globalization, it would also revolutionize travel. By today’s standards, Thomas Friedman’s “Flat Earth” would look annoyingly bumpy.
In the meantime, HPC will continue to be used for such tasks as saving the Earth from global warming, protecting the nuclear arsenal, and cracking the genetic code. But after that, the real fun begins.
As always, comments about HPCwire are welcomed and encouraged. Write to me, Michael Feldman, at [email protected].