News & Analysis
Sun picks optical links for advanced supercomputing
Loring Wirbel
11/14/2005 1:00 AM EST
Sun and Luxtera will demonstrate a CMOS photonic link for their Hero system proposal using four-wave dense wavelength-division multiplexing (DWDM) for a bidirectional 40-Gbit/second link. Mike Vildibill, director of HPCS product planning at Sun, said the Luxtera optical links are not intended to replace the chip-to-chip wireless interconnect that Sun developed for Darpa, called Proximity Communications. But it is not clear where the boundaries of the Proximity and Luxtera technologies will be drawn.
"Proximity is intended as a direct, bondless chip-to-chip interconnect, while the Luxtera optical links could be used between boards or between systems," Vildibill said. "But the precise relation between the two interconnect types is still being considered."
The intelligence agencies came to Darpa in 2002 with a plea to get computer vendors to think about overall performance in end applications, rather than raw hardware petaflops. Consequently, HPCS stands for "productivity, not just raw performance," Vildibill said. IBM Corp., Sun and Cray Inc. are engaged in a "bake-off," after which Darpa will anoint one to move to Phase 3 HPCS development before the end of fiscal 2006.
Sun's Hero project will combine interconnect technologies, advanced processor technologies (such as the multicore "Rock" Sparc project) and advanced server architectures to promote a scalable, reliable architecture to HPCS.
Luxtera CEO Alex Dickinson said that Sun and Luxtera already worked together on a $12 million Darpa project called Epic, for Electro-Photonic Integrated Circuit, in which the two joined with Freescale Semiconductor Inc. and other vendors to demonstrate parallel DWDM optical interconnect. Sun and Luxtera are extending this work to the system level by collaborating on Hero.
Because Darpa needs a reliable and manufacturable interconnect solution, an optical interconnect based on continuous-wave, rather than externally modulated, lasers was an important factor. Vildibill said that Sun carefully considered every technology from vertical-cavity surface-emitting laser arrays to tunable lasers, but decided that Luxtera's CMOS silicon integration of ring modulators, filters and lasers was the way to go. Even if Sun does not move on to HPCS Phase 3, its work with Luxtera will continue, the company said.
Luxtera's Dickinson said the Darpa work will help drive the direction of server-cluster components Luxtera will offer in the future. Commercial products planned for the 2006-07 time frame will be aimed at practical CMOS integration more than optocomponents, he said.



