The recent article titled, “Silicon Photonics is Hot” authored by Rick Merritt and based on an interview with Andy Bechtolsheim, spurred me to take a deeper look at the technology challenges we’re facing with silicon photonics as we scale switching networks from 10G to 100G and eventually one terabits. It’s a brave new world when we eliminate the tried and true hand assembly of hundreds of optical piece parts and move to silicon photonics. Some optical functions are easy to implement in silicon, while others are more difficult. We cannot, and should not, mix some of the old parts with a silicon chip to create a partial silicon photonics solution. Rather, we need the whole Optical Engine to be integrated onto the silicon platform.
The Optical Engine takes multiple high-speed electrical channels, converts them to optical, combines the channels to send them via a single optical fiber to any distance from as close as the next rack to as far as across the entire data center. At the receive end, the Optical Engine separates the light stream into separate channels and converts them back to electrical channels. In data centers, Optical Engines are emerging as the lowest power, smallest technology choice in pluggable transceivers for connecting cluster switches and routers and are used in Active Optical Cables to connect server blades and switches. In addition, Optical Engines will soon be embedded mid-board to reduce the power and increase the density in board-to-board applications.
However, there are many challenges in integrating optical functions on a CMOS platform designed primarily for electrical functions. Let’s have a look at each of the key opto-electrical functions and the challenges of full integration in a CMOS platform:
Lasers: Lasers provide the source of light for the Optical Engine, but some data center lasers are expensive. Kotura has developed on-chip functions that use the light from low-cost, low-speed lasers. The laser is the one optical component that is not monolithically integrated, but recent developments in flip-chip bonding of lasers and arrays make this a high-volume, low-cost process. On chip features eliminate the need for any lenses, isolators and beam collimators traditionally used in laser subassembly. Kotura’s laser design eliminates the need for expensive hermetic packaging. It just takes a few seconds on an automated station to align and bond a laser array to the silicon photonics chip. Finally, one of the toughest problems, getting a low-cost light source into the chip, has been solved.
The real value in optical networking is the ability to combine multiple wavelengths of light into the same physical pipe. For 100G interconnect, we use this dimension of parallelism, called Wavelength Division Multiplexing (WDM), to put four wavelengths of light onto a single fiber. Of course, four parallel fibers would also work, but that increases the cost of the network and wastes the bandwidth of the fiber. WDM allows us to scale, using the same data center fabric, to many more channels in the future.
WDM can be tough to do in silicon photonics, as both wavelength specific lasers and multiplexers are needed. However, we do not want to use expensive, wavelength specific lasers that are used in telecom networks. A better solution is to use generic lasers and convert them to wavelength specific ones by integrating a grating mirror on the silicon chip. By varying the location of this mirror, Kotura transforms each gain chip into a unique wavelength specific laser.