When Supercomputers Become Too Super

The power requirements of supercomputer centres are escalating beyond control, but new technologies may help.

For 30 years supercomputers surfed the edge of a technological wave. Each successive generation of processors used ever smaller integrated circuits. The energy to power each circuit went down with this shrinking, and despite an increase in the total number of circuits, the overall power requirements remained the same. The result: more computing capacity for the same electrical power.

But in the last five years this cycle has been broken. Circuits are still getting smaller, but the power they need is staying the same. So if you increase the computing speed of a computer by adding more circuits, the power requirement has to increase. We’re now seeing unsustainable growth in both the carbon footprint and power requirements of supercomputing centres.

If we’re going to build a new generation of supercomputers to help answer critical questions about climate change, molecules, or nuclear fusion, new, more energy efficient, designs are desperately needed.

Just how much power are we talking about? Let’s put things in perspective first. Power is measured in Watts – that’s the rating on a light bulb. Old fashioned incandescent light bulbs used to average about 60 Watts. A typical home with four people in North America needs, averaged over a day, about a thousand watts or a kilowatt of power.

About fifteen years ago, supercomputer centres required tens or perhaps hundreds of kilowatts. Fast forward to the present and that number has grown to between 20 and 30 megawatts. That’s a small city. By 2012, the largest installations will need over 100 megawatts. The annual power cost could soon dwarf the cost of the hardware.

Google was the first to come face to face with this problem. Their simple search box is driven by millions of computer processors across the world. Google has even overtaken governments when it comes to building massive computing infrastructure.

Note that Google isn’t quite doing supercomputing. While the distinction is somewhat technical, it can be summarized simply: Google wants to answer a huge number of small questions, while supercomputing centres tend to answer a small number of really big questions. Of course far more people want to use Google, and the implications of this warrants an article to itself.

Google is also ahead in the energy efficiency game. They moved their computers to places with cheap, renewable power. The data centre designs were radically overhauled to provide more efficient heating and cooling. Individual computers were chosen for power efficiency. All these improvements are now filtering through to supercomputer centres.

But a 50 per cent improvement here and there, through better cooling or something similar, won’t save you if other needs are growing by factors of 300 per cent.

So, electronic engineers are hard at work trying to design more efficient circuits. There are actually fundamental limits to how efficient circuits can be made, and the circuits we are building now, that power your Intel Core processor, for example, are surprisingly close to these limits.

Designers need a new approach. The most popular solution amounts to building the computing equivalent of a Prius rather than a Dodge Viper.

When it comes to cars, the more powerful it is, the less fuel efficient it is. It’s the same for processors. You can design a slower processor that is vastly more energy efficient. But how does that help? We want more computing power, not less.

You add speed by simply putting more processors into the computer. The increased efficiency from running slower more than offsets the increased number of processors you need. Of course, this all assumes that you can coordinate all these processors to work on your problem simultaneously, a concept known as parallel computing. It’s the computing equivalent of conducting symphony orchestra.

All current supercomputers are parallel, and the largest computers have over 200,000 processors working in tandem. What’s the problem with taking this still further? The answer is the sheer scale of the parallelism envisaged. Rather than dividing up into 200,000 pieces, circuit designers are now asking whether we could divide up problems into tens of billions of pieces.

This is going to be tough. Programming the current generation of supercomputers is difficult enough. Developing codes is slow and arduous, and commentators have dubbed this issue a “productivity crisis.” Imagine building a car that could go unimaginably fast, but you just couldn’t control it at those speeds. The next generation of supercomputers might face this problem.

Hence, supercomputer vendors like Intel, Cray, and Sun Microsystems are now funding computer language research with their own dollars. Unfortunately, years of academic research in this area has produced few practical advances. It’s hard to predict what will happen. But the stakes are very high, because even if we can build these machines, if we can’t program them, they won’t get sold.