Deal brings optical connections to Sun supercomputer

The server specialist will use technology from Luxtera, a 40-person start-up based in Carlsbad, Calif., to connect chips directly via optical links, the companies plan to announce Monday. The alliance will use Luxtera's silicon ring modulator, which chops up light into photons that can carry data and which can be integrated into silicon chips.

Sun already is employing one novel technology, the proximity communication, in a supercomputer design underwritten by the U.S. Department of Defense. The proximity interconnect lets overlapping chips communicate directly. That feature, combined with Luxtera's longer-distance optical connections, raises the possibility of building a computer with no electrical wires except the tiny ones within the chips themselves.

"Proximity communication brings high-bandwidth, low-latency data transfer between chips," said Mike Vildibill, the director of product planning for the Sun supercomputer. The technology partnership means that many more computing components can now be connected, he added. "With Luxtera, we're extending that incredible bandwidth to the module-to-module and rack-to-rack realms," he said.

Although a Sun computer using the Luxtera links won't be ready for years, the deal will be seen as a significant step toward the adoption of blending optical and silicon technologies. The companies plan to demonstrate a simple version of their technology at the SC05 supercomputing show in Seattle this week.

Optical networks have been around for decades, but to connect large computers over long distances. Squeezing fast optical systems to connect chips inside a smaller computer is a daunting task, because optical parts can't be shrunk easily. Optical networks transmit data using photons. The photons have to be created with a laser and then channeled into fiber-optic cables. Then, once received, the data has to be processed.

Typically, computers carry signals over wires using electrons, which are slower than photons and generate far more heat.

Combining the cheap techniques for producing silicon and wire chips with fast optical technology has been a dream of Luxtera, but also IBM and Intel.

Although the first Luxtera products will transfer data at 10 gigabits per second--the top speed of conventional networks today--the technology demonstration at SC05 will be of next-generation technology reaching 40Gbps, said Ashok Krishnamoorthy, a Sun distinguished engineer.

The higher speed is attained by sending four wavelengths--or colors--of light through the same fiber optic cable, a technology known as dense wave division multiplexing (DWDM).

The companies didn't disclose terms of their deal. But it could be important for Luxtera: Sun will start using the optical interconnect technology in supercomputers, Vildibill said, but then it will "trickle down" into conventional business servers. That's the bread and butter of Santa Clara, Calif.-based Sun.

Luxtera envisions future versions of its technology allowing data transfer on a single fiber-optic strand to reach 100 gigabits per second, then 1 terabit per second, and eventually 10 terabits per second. The company is developing technology now that permits 10 wavelengths of light to be transferred simultaneously. (Each wavelength corresponds to a specific frequency and color of light.)

Luxtera will start shipping its first sample chips to customers in February 2006, said Gabriele Sartori, vice president of marketing. Those models will use a single wavelength of light and will go into production a year later, he said.

Defense work
For now, Sun's goal is to build the Defense Department-funded technology prototypes. The work, to be completed by the end of 2006, is phase two of a plan called the High Performance Computing Systems program administered by the Defense Advanced Research Projects Agency.

Sun, Cray and IBM are vying to make it to phase three, the goal of which is to actually build a machine based on the prototypes. More than one company could be selected. Phase three had been expected to be done by 2009 or 2010, but now the machines it will fund are due to arrive by 2011.

Phase two will incorporate proximity communications and Luxtera's optoelectronic technology, but most likely won't yield a full-fledged computer, Vildibill said. "I don't believe we'll have a fully functioning computer system prototype by end of phase two," he said.

Proximity communication itself could well arrive by 2008 or 2009, Sun has said.

Sun hopes its DARPA-funded machine will swing the supercomputer innovation pendulum back toward mammoth machines and away from "cluster" systems, in which many small, independent machines are connected with high-speed networks and a powerful switch. The ultimate hope is that data transfers across the system will have as close to the same speeds and delays as local chip-to-chip communications, Vildibill said.

"Ideally you'd want to see a one-to-one ratio," Vildibill said. "The trend in the industry is going the wrong way."