University of Wisconsin-Madison engineers have designed and built the world’s most energy-efficient supercomputer, earning the top ranking on an important benchmark called the Green Graph 500 list.
Topping the Green Graph 500 list is an achievement that establishes UW-Madison as a major player in the field of high-performance computing. The certification builds on established practices from the U.S. Environmental Protection Agency Energy Star, and is backed by a steering committee composed of more than 50 international experts from academia, industry and national laboratories.
Developed as a complement to the Graph 500, which is a benchmark to evaluate performance for one of the most difficult problems in data-intensive computing, the Green Graph 500 measures how much electricity a supercomputer consumes as it grinds through the immensely challenging set of calculations. First place goes to the computer that runs the programs using the least amount of power.
“We are very proud,” says Jing Li, an assistant professor of electrical and computer engineering at UW-Madison who led the supercomputer’s creation. “Historically, the top has always been IBM or national labs.”
Achieving greater computational power per watt of electricity consumed will help enable faster computers with smaller ecological footprints—a win-win for human users and the planet.
Most supercomputers gobble tremendous amounts of electricity to perform their calculations. For example, “Summit” at Oak Ridge National Laboratory—the most powerful machine in the world—peaks out at roughly the same amount of power consumption as 7,000 homes.
Supercomputers excel at running conventional, linear programs. But there’s another class of problems that is far more challenging. Those are called graphing algorithms.
“Graphs are very difficult for a traditional computer,” says Li. “All the general tricks we know don’t work well.”
Traditional computers struggle with graphs because the data isn’t arranged in a predictable, linear pattern. Yet graphs are the best way to represent numerous important phenomena for our daily lives—for example, interpersonal connections among a social network, the electricity grid for a major city, or the interstate highway system.
Because there are several possible paths to connect individual nodes (think six degrees of Kevin Bacon), it’s nearly impossible to predict the most efficient way for a computer to map out a graph. Most computers end up wasting processing power by scattering closely connected nodes in far-flung locations across their memory.
That’s why Li and her student, Jialiang Zhang, took a unique approach—and created a computer unlike anything else in the world.
“The components are new, the ways we hooked things up are new, and the ways we mapped the software onto the hardware are new,” says Li.
Using advanced mathematical techniques, Li and Zhang devised a new method of data analysis that allowed their computer to store and map graphs more efficiently. Then, taking a top-down approach, they built hardware that could handle the demanding software needs.
They’re currently working to scale up the system to tackle larger and even more complex datasets than the Green Graph 500.
Key to their approach is a holistic mindset, bridging the gap between the two traditionally disparate fields of computer engineering and computer science by building both hardware and software together to be perfectly suited for the task at hand.
“If you want the computer to work more smartly,” says Li, “you have to make the hardware and software coherent in one system.”
Author: Sam Million-Weaver