A Smarter Supercomputer

A new design could run ultrahigh-resolution climate models.

Despite all the horsepower of today's supercomputers, none are powerful enough to run algorithms that predict future weather down to the kilometer--a resolution that would allow researchers to model cloud behavior, improving overall accuracy of climate models and enabling better policy decisions on global warming. But now, researchers at Lawrence Berkeley National Laboratory, in Berkeley, CA, have designed a new type of supercomputer that could be powerful enough to run models with kilometer-scale precision. In addition, the proposed supercomputer, which leverages chip design technology from cell phones and MP3 players, would be hundreds of times more power efficient than any other supercomputer, making it more cost effective to run.

Test run: Researchers simulate the layout of one of 20 million processors that will be used to build a power-efficient supercomputer that can model climate change with kilometer resolution. Credit: Kate Greene

Multimedia   See scientists discussing a new approach to energy-efficient supercomputing.

"Right now with supercomputers, money is the obstacle to building a bigger system to tackle bigger problems," says John Shalf, a computer scientist at Lawrence Berkeley National Laboratory. Using today's technology, a supercomputer capable of running high-resolution climate models would produce an energy bill of about $150 million a year, Shalf says. But by building a supercomputer with processors that are fine-tuned to save power, similar to the approach used in the mobile electronics industry, Shalf and his coworkers expect to be able to run a supercomputer for a fraction of the cost.

The Berkeley supercomputer is part of an effort by climate scientists to bring more computational power to highly complex climate models. Earlier this month, the National Center for Atmospheric Research, in Boulder, CO, announced that it will use an IBM supercomputer to look at the effects of climate change. The Berkeley supercomputer would, however, be a vast improvement because it could be far more energy efficient.

General-purpose processors, such as those in personal computers, have become powerful and cheap enough that many researchers simply wired up clusters of these standard processors to run intense computational tasks. While this approach works for a number of scientific problems, it fails for high-resolution climate models: the computation simply requires too much power.

Story continues below

Whereas a general-purpose processor is built to handle a range of operations, an iPod processor, for instance, is specialized for only a small number of tasks. The Berkeley researchers used software from Tensilica, a chip manufacturer in Santa Clara, CA, that develops chips for Motorola, to build a core with only the necessary functions. In addition, the researchers custom-designed the chip's memory and communication structure between cores to reduce inefficiencies, minimizing power consumption.

In total, the supercomputer will consist of 20 million processing cores. The researchers estimate that, with 32 cores on a chip about the size of a matchbook, the entire supercomputer will fill a space the size of half a tennis court. By the end of the year, the researchers expect to be able to simulate a climate model on tens of cores. This approach will allow the researchers to change the hardware to better fit the model, as well as modify the model to fit the hardware. "Instead of building a computer and then throwing a science problem to it, we're looking at the science problem and building a computer to fit the problem," says Shalf.