You might be unaware that Intel makes photonic modules for optical compute interconnect for datacenters, but the company has been in this business for several years. Intel Labs has been investigating ways to reduce the cost of these modules using automated semiconductor manufacturing techniques since the early part of this century. Back in 2006, Intel and the University of California Santa Barbara (UCSB) announced construction of the world’s first electrically powered Hybrid Silicon Laser using standard silicon manufacturing processes. That laser melded the light-emitting properties of Indium Phosphide (InP) with the light-routing capabilities of silicon by fusing a sliver of InP onto a silicon wafer, creating a single hybrid semiconductor laser chip. Now, Intel Labs has built upon this work by announcing successful construction of an eight-wavelength, distributed feedback (DFB) laser array on a chip that delivers excellent output power and wavelength spacing uniformity that exceed industry specifications for optical laser communications modules. This semiconductor component is the essential element in a wavelength division multiplexed (WDM) laser for high-speed optical communications within datacenters.
A WDM laser array generates multiple beams of light at different frequencies that can be combined onto one fiber, thus pushing more bandwidth through the optical fiber than achieved by a single-frequency laser. In the recent announcement from Intel, the experimental eight-laser DFB array generates eight beams of infrared light with uniform 200GHz channel spacing. As with the Hybrid Silicon Laser announced in 2006, the DFB laser array melds the light-emitting properties of Indium Phosphide (InP) with the light-manipulating capabilities of silicon, but in a far more sophisticated manner.
First, optical gratings etched into silicon sit adjacent to each InP light emitter, tuning the light from each emitter. The physical pitch of each of the eight gratings causes each laser to generate a precise beam of light at a different frequency. These gratings are made on Intel’s standard 300mm CMOS wafer line using deep-UV immersion lithography, which is a high-volume, low-cost process. Light pipes, modulators, combiners, and detectors are also etched into the silicon using the same high-volume CMOS process technology. V grooves cut into the silicon using anisotropic etching serve to precisely align optical fibers to the hybrid photonic chip. All of the critical components of a high-speed optical transceiver are made simultaneously, on what amounts to an optical integrated circuit.
The ability to manufacture the most critical photonic components as a hybrid, monolithic chip will allow Intel to greatly improve the bandwidth of its optical modules while reducing their costs relative to today’s modules. Haisheng Rong, senior principal engineer at Intel Labs, said that this technique could be extended beyond eight-laser arrays to create photonic components that generate, modulate, and combine 16 or 32 wavelengths simultaneously, so there’s plenty of headroom in this technology.
Beyond the development of new, faster optical communications modules for use in datacenters, these semiconductor lasers are particularly amenable to Intel’s chiplet manufacturing approach where different semiconductor die are packaged together to create a vastly more complex device. Viewed through the appropriate lens, chiplet manufacturing is one way to extend the life of Moore’s Law.
The development of single-chip, DFB laser arrays as chiplets enables co-packaged optics. That means Intel could develop CPUs, GPUs, FPGAs, memory modules, and storage devices that all communicate optically instead of using copper wires. At current communication rates that exceed 100Gbps, optical communication is more energy efficient than electrical communication, so the technology holds promise for revolutionary breakthroughs in chip-to-chip interconnect.