Supercomputers for Rent
About 15 years ago, supercomputers were thought of as rare and exotic creatures. Government laboratories in the U.S. and Japan spent hundreds of millions of dollars on custom computing rigs and specialized facilities to house them in a bid to tackle the world’s toughest problems. When a new supercomputer came to life for the first time, the research center would hold a press conference, inviting onlookers to witness its calculating prowess.
And today? Well, you can sit down at the kitchen table, whip out a credit card, and rent the same amount of horsepower by the hour online. The master of the supercomputer rental is Jason Stowe, the 35-year-old chief executive officer of Cycle Computing based in Greenwich, Conn. Stowe’s company has written software that can talk to Amazon.com’s cloud-computing service and coordinate 50,000 or more of its computers to work on a single problem. Through a couple of years of trial and error, Cycle Computing has gotten the rental rate for these number-crunching beasts down to about $1,000 per hour. “This really is the democratization of large-scale computing,” Stowe says. “Now, even someone with a modest research grant can work on solving the most challenging questions.”
Supercomputers, like most computers, for years have been falling in price while gaining processing muscle. It’s now quite common for a large company to own a supercomputer-class machine that’s built by lashing together thousands of standard computers. Still, such a machine costs more than $20 million to build and a few million more each year in power and cooling costs. And employees often have to fight for time on the machine, since there’s such a high demand for this limited resource.
Stowe, a math whiz-turned-entrepreneur, realized that Amazon’s cloud computing had gotten good enough to change this equation. He and his company’s engineers developed software that scours up to seven of the data centers Amazon has scattered across the globe for large numbers of free machines. Cycle Computing chops up computing problems into small chunks that can be spread across all of the machines and then reassembles the results into an answer. The company deals with any glitches that arise when an Amazon computer dies in mid-calculation. “Amazon is wonderful, but it still does have errors,” Stowe says. “You have to correct for those in a compressed time frame.”
“Cycle has done a great job demonstrating what’s possible,” says Adam Selipsky, vice president of Amazon Web Services. A number of companies have followed Cycle’s lead and gone on to build their own supercomputer-class systems, according to Amazon. The first supercomputer rental-service clients were large banks, insurance companies, and engineering firms, Stowe says. All of the companies had their own large computing systems but were looking for occasional access to more horsepower for tough calculations they needed to complete in a hurry. Jobs that might take six weeks could be done in 18 hours.
One company pushing the rental limits is Schrödinger, a pharmaceutical technology firm that uses software to aid in drug discovery. The company runs simulations to see how molecules—drugs and their targets—will interact. Schrödinger has its own high-powered computer and can spend 275 hours going through the permutations for one simulation. With the aid of Cycle Computing and Amazon, though, it did the same work in 3 hours for $4,828 per hour. “It’s amazing,” says Ramy Farid, president at Schrödinger. “I think everyone can look at the trends and knows that this price will just keep going down.” (The price per hour today fluctuates based on an auction process run by Amazon.)
Cycle Computing’s customers include heavyweights such as JPMorgan Chase, Pfizer, and Lockheed Martin. The 20-person company has yet to take any funding but may seek it out now that the rental business has started to boom, Stowe says. The hope is that smaller companies will take advantage of the technology as well, using a supercomputer rental to put them on equal footing with more established competitors. “A startup doing a new product design can now get a system as good as that at Procter & Gamble or General Motors in only a few hours,” Stowe says. “This opens up a whole new path for innovation.”