Supercomputers run the vast simulations that help us to better predict climate change--but they also contribute to it through their energy consumption.
The lifetime cost of powering supercomputers and data centers is now surpassing the cost of buying the machines in the first place. Computers, small, medium, and large, have become a significant fraction of energy consumption in developed countries. And although the authors of Superfreakonomics may not understand it, the carbon dioxide used to supply this energy will absorb, during its time in the atmosphere, some 100,000 times more heat energy than that.
To draw attention to this issue, for the last two years researchers at Virginia Tech have been reordering the Top500 list of the fastest supercomputers, ranking them according their energy efficiency in the Top Green500 list. I have a wee news story out today on the newest list, released last Thursday, on the web site of the Communications of the Association for Computing Machinery. The top-ranked systems are from the QPACE project in Germany, and are designed for quantum chromodynamics calculations.
Calculating efficiency isn't as straightforward as it sounds. The most obvious metric is the number of operations you get for a certain amount of energy. This is essentially what Green500 measures in its MFLOPS/W, since MFLOPS is millions of floating-point operations per second and watts is joules per second.
As a rule, however, this metric favors smaller systems. It also favors slower operation, which is not what people want from their supercomputers. Some of the performance lost by running slowly can be recovered by doing many operations in parallel, but this requires more hardware. For these reasons, the most efficient systems aren't supercomputers at all. The Green500 list works because they only include the powerhouse machines from the Top500 list, which puts a floor on how slowly the competing machines can go.
Over the years, researchers have explored a family of other metrics, where the energy per operation is multiplied by some power of the delay per operation: EDn. But although these measures may approximately capture the real tradeoffs that systems designers make, none has the compelling simplicity of the MFLOPS/W metric. This measure also leverages the fact that supercomputer makers already measure the computational power to get on the Top500 list, so all they need to do extra is measure the electrical power in a prescribed way.
These systems derive much of their energy efficiency from the processor chips they use. The top systems in the current list all use a special version of IBM's cell processor, for example. I worked on power reduction in integrated circuits more than a decade ago--an eternity in an industry governed by Moore's Law--and some of my work appeared in a talk at the 1995 International Electron Devices Meeting. I also served for several years on the organizing committee of the International Symposium on Low Power Electronics and Design, but I'm sure the issues have advanced a lot since then.
In addition to the chips, the overall system design makes a big difference. The QPACE machine, for example, serves up as about half again as many MFLOPS/W as its closest competitor by using novel water-cooling techniques and fine-tuning the processor voltages, among other things. These improvements aren't driven just by ecological awareness, but by economics.
There's still lots of room for improvement in the energy efficiency of computers. I expect that the techniques developed for these Cadillac systems will end up helping much more common servers to do their job with less energy.