create an index that actually gives you some measurement?” Just as you’d measure an automobile’s fuel efficiency by looking at its miles-per-gallon rating rather than its top speed, Stone argues, computer buyers should consider how much bang they’re getting for each buck they spend on electricity.
SiCortex’s index isn’t the first to rate computers on their green credentials. Wu-chun Feng, a computer scientist at Virginia Tech, introduced the Green500 list last November; it essentially re-orders the computers on the Top500 list according to their energy efficiency in megaflop/s per watt. But SiCortex argues that the Green500 list is inadequate, since the Top500 performance statistics are based solely from the Linpack benchmark, which tends to reward CPU speed over other factors such as the speed of memory subsystems and of the communications network connecting processors. As I explained in a September profile, SiCortex’s machines get most of their speed boost from their backplanes, the communications meshes linking all their processors—explaining why the company is basing its proposed index on the HPC Challenge, a broad collection of seven benchmarks devised by the Defense Advanced Research Projects to measure computers’ performance in real-world situations.
You probably won’t be surprised to learn that SiCortex computers compare very well to competitors’ machines in statistics distributed by the company today. In one chart, SiCortex’s 1,458-processor machine bested not only the Swiss Cray XT3 on the Linpack component of the Green Computing Performance Index, but also beat an IBM Blue Gene computer at the San Diego Supercomputer Center (136 gigaflop/s per kilowatt), an SGI Altix ICE 8200EX system a the Dresden University of Technology (232 gigaflop/s per kilowatt) and a Cray XT4 system at the U.S. Army Corp of Engineers’ Major Shared Resource Center (130 gigaflop/s per kilowatt).
“We obviously come out pretty well, because of the way we designed our machines,” says Stone. “But there are cases where we don’t come out on top. And maybe somebody will produce data that make us look terrible. That’s fine—we’re providing everybody with the data we used and where we got, so that they can go off and find what’s wrong.”
And as it turns out, SiCortex’s machines are not the greenest around. When IBM’s Roadrunner is running at 1.026 petaflop/s, for example, it consumes 2,346 kilowatts of power, which comes to an extremely efficient 437 gigaflop/s per kilowatt, earning it the #3 spot in the Green500 list. That’s almost twice as efficient as a 1,458-processor SiCortex machine. Of course, 2,346 kilowatts is an immense amount of power—equal to the output of two or three nuclear power plants—and not too many organizations can afford Roadrunner’s $100 million price tag.
Stone says SiCortex hopes to find an neutral, outside organization to take over administration of the Green Computing Performance Index—and mentions Jack Dongarra, the creator of the Top500 list, as a candidate. “We’d be happy to work with him,” says Stone. “Or if a new body of people wanted to get together and formulate a new index around this data, but maybe call it something different or measure in a different way, we’d be happy to go along with that. Debate is healthy, and we need a good debate about how to measure the computing industry’s carbon footprint.”
Addendum 11/7/08: HPCwire’s John West published a nice feature yesterday on SiCortex and the Green Computing Performance Index.
