There’s no such thing as a free flop.
A “flop,” or floating-point operation, is a calculation involving a decimal number; engineers often measure the performance of computers in terms of the number of “gigaflop/s” (billions of floating-point operations per second) they can sustain. In the supercomputing community, it’s a longstanding ritual to compare various machines’ maximum gigaflop/s as they tackle standardized math problems called benchmarks. The world’s fastest computer—a 122,400-processor IBM machine at the Los Alamos National Laboratory called Roadrunner—can run at just over a million gigaflop/s, or 1.026 petaflop/s, according to Top500, the most prominent list of top supercomputing sites.
But as a general rule, the faster a computer runs, the more power it consumes—and the more waste heat it generates, and the more additional power is needed to run cooling systems. If current trends continue, according to McKinsey & Company, then by 2020 the electric plants needed to power the world’s data centers will be churning out more greenhouse gases than the entire airline industry. The problem is getting so serious that some organizations are having to scale back plans to upgrade their data centers with faster machines—not because they can’t afford them, but because local utilities can’t supply any more electricity.
Maynard, MA-based SiCortex serves the high-performance computing market, so it’s naturally obsessed with the gigaflops game—but with a twist. The six-year-old startup builds massively parallel computers with thousands of processors. The processors themselves aren’t very fast. They run at around 700 Megahertz, slower than the chips inside most desktop and laptop PCs, which saves a lot of electricity. But they’re wired together in a way that makes SiCortex’s computers extremely zippy nonetheless. And today SiCortex is proposing an overhaul in the way the performance of high-end computers is measured and ranked, one that would take a machine’s power consumption into account and reward machines that use it sparingly.
The company calls its new measuring system the Green Computing Performance Index, and it’s urging managers of government and academic supercomputing centers and corporate data centers to use it to evaluate the full benefits and costs of owning high-performance computing systems from companies like Cray, SGI, Hewlett-Packard, IBM, and of course, SiCortex itself.
When evaluating the full cost of owning a high-performance computer, SiCortex argues, organizations should divide its performance in gigaflop/s by its power consumption in kilowatts. When you do that, a number of machines that are nominally faster than SiCortex’s machines—such Cray Inc.’s XT3 and XT4, IBM’s Blue Gene, and SGI’s Altix 8200EX—come out looking like power hogs. When running the standard Linpack benchmark, for example, an 1,100-processr Cray XT3 machine at the Swiss National Supercomputing Centre gets just 17 gigaflop/s to the kilowatt, while a 1,458-processor SiCortex machine gets a comparatively huge 253 gigaflop/s to the kilowatt.
“If you look at the high-performance computing benchmarks that Top500 produces, it’s great stuff, but it doesn’t give you a measure of the actual energy efficiency of the computers themselves,” says Christopher Stone, SiCortex’s CEO. “Everyone in the high-performance computing business is running around talking about being green or wanting to be green, so we thought why not
