Tianhe-1A: Computnik Revisited?

China has just unveiled the world's fastest supercomputer. Will the U.S. fight back?

Yesterday, China unveiled the world's fastest supercomputer. The Tianhe-1A system is powered by 7,168 graphics processors and 14,336 Intel Xeon processors. It bests the previous top performer, the U.S. Department of Energy's Oak Ridge National Laboratory “Jaguar” system, by over 40 per cent. Twitter was abuzz with “Woohoo! We're number two!” comments. Does this mark the end of U.S. hegemony in supercomputing?

These concerns have been heard before. In 2002, Japan took top honours with the “Earth Simulator” supercomputer. The American press filled page after page with commentary on how the U.S. was falling behind. Harkening back to the days of the Cold War and “Sputnik,” the media coined the “computnik” moniker, and it quickly became the word du jour in computing circles. Coming right in the depths of the dot-com crash, Japan's supremacy was especially painful for the U.S.

Yet the U.S. responded by simply going out, spending more money, and building something bigger and faster. Having the fastest supercomputer is a lot like having the tallest skyscraper – by the time you finish yours, another, bigger one is months away from completion.

Supercomputing is no longer about blue-skies or weapons research. It's a critical part of mainstream business, like oil and gas exploration, financial analysis, and industrial design. The techniques developed on the very largest research systems are rapidly deployed elsewhere. Indeed, one can chart a straight line between advances in supercomputing made in the mid 1990s and the facilities now installed by companies across the globe.

Almost 10 years on from the “computnik” event, the international competition is quite different. Despite plans in Japan to build a much larger system than the Earth Simulator, the “Keisoku” project, economic woes appear to have put it in the slow lane. But unlike Japan’s, China's economy is not seriously faltering. China’s investments in supercomputing are making everyone take notice. In the June 2010 TOP500 list of the world’s fastest supercomputers, China moved to second place behind the U.S. in total installed capacity.

Tianhe-1A is actually relatively cheap. At US $88 million, it undercuts the reported cost for the Jaguar system by $20 million. But numbers like these need to be taken with a pinch of salt. Manufacturers frequently offer deep discounts to win these enormous contracts and in some cases actually lose money. Nonetheless, taking these numbers at face value, Tianhe-1A looks like great value for money.

But Tianhe-1A is not your typical supercomputer. Most of the supercomputers on the TOP500 list are built by connecting together thousands of regular CPUs from either Intel or AMD. Instead of CPUs, Tianhe-1A uses graphics processors to give it a speed advantage. But this isn't a magic bullet for faster supercomputers. When it comes to programming, Tianhe-1A isn't nearly as flexible as Jaguar. It's a bit like comparing a Porsche with a souped-up SUV. One is faster than the other but one can go places the other can't.

But whether or not that is an issue depends entirely on what you want to do. If you have problems that work well on graphics processors (such as weather forecasting), then this is the architecture to buy. Programming graphics processors is still harder than regular CPUs, but tools are getting better all the time and programmers are rapidly becoming more familiar with them. Once you have code that works on graphics processors, or special-purpose hardware in general, they really make sense.

And this looks to be the future of computing in general. Rather than one very general-purpose processor that can do just about everything, we're seeing an evolution towards special-purpose processors that can do certain tasks extremely well. Due out next year, AMD's Fusion processors, which combine the graphics processor and CPU, are perhaps the first mainstream step in this direction.

Perhaps in the future, the distinction might seem irrelevant. Early generations of Intel CPUs include co-processors for speeding up floating-point calculations (i.e. those involving decimal parts). Starting with the 486 processor, that functionality was included on the CPU. It's hard to imagine a CPU without floating-point units today. But when you run certain types of code, those floating-point units just sit idle. One of the key goals in future chip design will be to make sure the idle units don't consume too much power.

As China reaches the peak of the supercomputing world, are we witnessing the true beginning of Chinese hegemony? Or in these troubled economic times, will the U.S. fight to get back on top?