IBM researchers worked several angles to wrings costs from a 160-Gigabit-per second optical transceiver module that will be presented Thursday (March 29th) at the Optical Fiber Conference being held in Anaheim, Calif. this week.
The “Terabus” project at IBM’s Yorktown Heights lab employed plain-vanilla 130-nm CMOS, rather than silicon-germanium BiCMOS, for the 16-channel full-duplex (transmit and receive) transceiver IC. Hitting 160-Gbps in a 130-nm CMOS process was an accomplishment itself. The greater challenge was to develop manufacturable alignment and flip-chip packaging techniques to place the photodiode (on an Indium Phosphide substrate) and VCSEL components (in Gallium Arsenide) together with the CMOS component, all in a low-cost organic (FR-4) carrier.
“In hindsight it appears relatively simple,” said research staff member Fuad Doany. Designing the CMOS transceiver was relatively straightforward. Aligning the optoelectronic components, with integrated lenses, in an integrated device was the more difficult challenge.
“We didn’t want to do something exotic that couldn’t be done at a manufacturing level. Getting them into an integrated device was the hard part of the work. The intent was to put it in a system. That is what makes this powerful. In the end this looks like a typical IC bonded to an organic carrier,” said research staff member Clint Schow.
Marc Taubenblatt, senior manager of the project, said the project received support from Darpa, which was looking for a way to create optical circuit cards.
“What we are aiming at here is very low-cost optical transceivers that have a chip-like format,” said Taubenblatt. “We wanted to use a commercializable technology for replacement of copper traces, to compete with copper electrical interconnects (on cost) but with much higher performance.”
The original purpose was to provide much faster optical links between processors in high-performance computers. With the advent of video stressing the performance of home networks, IBM now envisions the day when the optical links could be cheap enough to connect a computer with a high-definition television in the home, for example.
The transceiver discussed this week is intended to work with polymer waveguides built into an organic substrate to do chip-to-chip interconnects on a card. The transceiver module would be positioned close to an MPU so that the signals do not travel far electrically before going to the optical module.
For long-haul connections greater than a kilometer in length, single-mode fiber is used. However, it is smaller in diameter and harder to align to than multi-mode fiber. Multi-mode fiber with low-cost transceiver modules would provide high-speed connections between circuit boards spaced perhaps a meter apart, or for connections of perhaps 100 meters on a corporate or academic campus, or within an apartment building or home network.
“We believe this can be made to compete with the standard Ethernet infrastructure out there,” Taubenblatt said. The Ethernet Working Group is developing a 100-Gigabit-per-second standard, expected by 2010. The IBM transceiver module will be compatible with the next-generation Ethernet standard, he said.