In recognition of the huge importance of graphics and gaming to the future of computing, Advanced Micro Devices (AMD), of Sunnyvale, Calif., is building the fastest commercial supercomputer in the world and selling its use to makers of online games. When it’s ready, in the second half of 2009, it will manage a thousand million million floating-point operations per second—a petaflop. That will put it on a par with Roadrunner, the U.S. Department of Energy’s most powerful machine.
The idea is to compute a game’s graphics, compress them, and send them out over the Internet so that online gamers can run the results on platforms, such as cellphones, that are too computationally puny to render the graphics on their own. Game makers would write their software for the supercomputer—rather than for a PC, smart phone, or other platform—and then rent computer time and bandwidth on AMD’s machine. It’s a particularly striking example of the shift in the balance of power away from the platform and into the network, or cloud—hence the computer’s name, the Fusion Render Cloud.
Cloud computing makes it ridiculously easy for potential customers to try out a game and get hooked—all they have to do is go to a Web site and start playing. Another plus is that by off-loading most of the computational guts, cloud computing puts game software out of the reach of pirates. And it economizes on computation time by supplying users with only as much graphical detail as the available bandwidth can handle, when it can handle it.
”We’ll be able to render 30 to 50 frames per second remotely and stream it to any modern Web browser,” says Jules Urbach, chief executive of Otoy International, the game technology company that’s writing the software. ”If you’re on a low-band phone, we’d send less data per frame—for a lower-res image—but you’d still get 30 to 50 frames per second.” The system tests the connection throughout a session, so if the user changes to a more capacious network—say, by switching from a cellphone to a hard-wired desktop—the system responds by adding detail.
The supercomputer will consist mainly of 1000 graphics-processing units (GPUs), each of which can run 800 concurrent computational tasks, or threads. At two chips per graphics card, that means 1600 threads at a pop. About a quarter of the chips will compress or decompress data, and the rest will render the graphics. That’s more than enough muscle, AMD says, to serve the players in any existing massively multiplayer game. The company says it can scale up the machine simply by adding more cards.
AMD says the Fusion Cloud will consume only 150 kilowatts running full tilt, compared to Roadrunner’s 2.35 megawatts. AMD says it will be able to perform as many operations per second as Roadrunner with less power, because its GPUs, with their hundreds of cores, can do far more things at once than can central-processing units (CPUs), which have only a handful of cores. That allows GPUs to better render graphics, a task best handled by dividing it up and working on the pieces in parallel.
The Fusion Cloud is happening now because three things fell into place only recently. Rapid-access memory reached the gigabyte range; a way of getting GPUs to perform floating-point operations—long the domain of CPUs—was found; and the rendering of subtle shades of light got much easier to program. ”Graphics cards originally had no shaders,” explains Urbach. ”Then they got geometry shading and then pixel shading, an enormous leap. Finally, since June of last year, we’ve had a general-purpose shader.”
AMD’s plan is merely a battle in a larger war being fought by Intel, AMD, and Nvidia Corp. over the advantages of graphics processing. ”We compete against Intel on the CPU side—they’re trying to get their CPUs to be more parallel—and against Nvidia on the GPU side—they’re trying to get their GPUs to be more flexible,” says AMD spokesman David Nalasco. ”We have both.”