Silicon photonics: Will the hare finally catch the tortoise?
I asked Sylvie Menezo, Director of France's IRT (Technological Research Institute) Nanoelectronic Silicon Photonics Program, the question I ask every time I talk to someone about silicon photonics. Most people waffle around questions of power, heat, and cost whenever lasers are involved. Menezo is the first person to give me a direct answer. My question: at what data rate do you think engineers will have to focus on designs that use photons to push data? Her answer: 50 Gbps per optical carrier for transmission farther than two meters.
By overcoming the signal degrading characteristics of electrical interconnects—loss, messy frequency response, crosstalk, and impedance matching difficulties—with clever technologies such as pre/de-emphasis, embedded clocking, and equalization, the need for silicon photonics has been pushed into the future, beyond 28 Gbps. Optical interconnects at 10-100 Gbits/s suffer chromatic and polarization-mode dispersion, which are tiny levels of loss and reflections, but only after propagating hundreds of meters. Optical eyes stay wide open and awake over reaches of dozens of meters.
The people who actually do silicon photonics are comfortable when it comes to including fiber-optic applications in their field. Being a fan in the cheap seats, however, I think it's fair to distinguish fiber optics from silicon photonics. A purist's silicon photonics is not infected with fibers, that's fiber optics. Genuine silicon photonics applications transmit photonic signals across waveguides within a chip, or chip-to-chip, or what have you, but not through transceivers coupled to fibers. Last year at the Intel Developer’s Forum they had a whole silicon photonics demo area but all I saw was fiber optics.
Menezo sent me a nice back-of-the-envelope summary of the state of silicon photonics: At 25 Gbps, electrical links are limited to about 5 m, generate about 6 mW/Gbps and cost about $0.20 per Gbps. VCSELs (vertical-cavity surface-emitting lasters) reach 20 times farther on multi-mode fibers, generate about 3 times more heat, and cost about ten times more. For mainstream apps, she said that they need to get the power down to 1 mW/Gbps and reach up to 1 km (on fiber) and be able to operate at a temperature of 80C (175F). She said, "The laser used to be an issue, but several teams have managed its integration on Si photonics at the wafer level." The remaining issues are moving the process to CMOS foundries and cost.
Samtec, the company famous for the DesignCon mEEt and gEEk parties, stuffed tiger swag, and interconnects, connectors, and doodads of all types, recently joined France's IRT Silicon Photonics Program as part of their push into silicon photonics. Marc Verdiell, CTO of Samtec's Optical Group, said that IRT is "a bit ahead in their thinking compared to most, and realize the importance of developing a cost effective, scalable, small size optical packaging solution, which is one of our important contributions to the consortium."
Which brings us to the cost issue. Verdiell said, "Cost of developing the chips is much higher than for regular transceivers" and isn't "amortized over volume enough to be competitive (yet)." The second cost issue, Verdiell said, is packaging: "Silicon Photonics does not have a cost and size competitive single'mode packaging solution (yet).” Remember, single mode, as opposed to multi-mode, fibers are the ones capable of transmitting over kilometers.
The path from electrical signaling to the purist's Si photonics will include intermediate steps. PCIe's OCuLink cabling combines optical and electrical connectors; with the laser and receiver integrated in the cable and powered at the connection, you get a clean electrical signal and don't need to worry about what happens between here and there. Samtec's Firefly (Figure 1) and Molex's Fireflex products include OCuLink applications, but they also have Flyover parts that get closer to pure Si photonics. The Flyover 14-28 Gb/s components integrate optics with fibers, for on-board and on-package, chip-to-chip designs; the data “flies over” the circuit board. To beat the heat, they use different types of heat sinks.
Figure 1. The Optical Firefly Flyover component with heat sink from Samtec (Source: Samtec).
Genuine silicon photonic systems without fibers transmit signals through waveguides integrated into the silicon. Because light doesn't behave nicely in pure silicon waveguides—due to nonlinearities like two-photon absorption, stimulated raman scattering, and the kerr effect—nicely behaving optical channels are made by depositing polymer waveguides onto the silicon that includes the laser, modulator, and receiver—usually an APD (avalanche photodiode). See Figure 2. These systems don't require independently packaged transceivers, connectors, and fibers, but they do face the developmental problem of coupling optical signals from pure silicon to polymer.
Figure 2. A CMOS silicon photonics chip. Credit: IBM Research. (Source: Optics.org)
Verdiell said that they're pushing forward. Still not quite targeting the purist's fiber-free Si photonics (yet), they're "developing a new generation of fiber backplane optical connectors that will allow much higher density and much reduced protrusion, as well as modularity, but still over standard fibers while the polymer waveguides solutions are developed."