SILICON PHOTONICS: On-chip router doesn’t muddle signals
We should all look forward to the day when optical interconnects become an integral part of computer chips.
We should all look forward to the day when optical interconnects become an integral part of computer chips. The reason? At some point soon, it will be the only way the burgeoning number of multiple processors in our personal computers will be able to talk fast enough to each other.
FIGURE 1. A simple 4 × 4 single-wavelength router has certain switching states in which two data paths (red and blue) share the same physical path (purple cross)—rendering the two data streams unable to be separated.
While practical intrachip optical communication is still years away, scientists at Cornell University (Ithaca, NY) and Columbia University (New York, NY) have developed an optical-waveguide-based single-wavelength router that could become an essential part of such systems.1 While other such routers have been created, the new router is “hitless”—meaning that two different data streams never share the same path.
Single-wavelength on-chip networks need only one light source, saving space and power compared to systems that route signals based on wavelength and thus require multiple sources. Tunable microring resonators can serve as routing elements in single-wavelength routers; however, a simple mesh configuration, for example in a 4 × 4 router, results in shared paths for some situations, meaning that the simple router’s practical usefulness is limited (see Fig. 1).
In these sorts of routers, two microring resonators are located at every intersection of two waveguides, one resonator for each incoming data signal. Tuning a resonator’s resonance away from the source wavelength allows an input data stream to flow unimpeded down a waveguide, while tuning the resonator’s resonance to the source wavelength shifts the data stream to the perpendicular waveguide.
To eliminate the potential shared paths that occur in the simple 4 × 4 router, the Cornell and Columbia researchers came up with a new geometry (see Fig. 2). While somewhat more complex, the new layout not only permits all routing options with no shared paths but requires no more rings than the simpler geometry.
The microring resonators, which have a 10 µm radius, are temperature-tuned by nichrome heaters located atop the rings. To fabricate the router, a standard silicon-on-insulator wafer with a 3 µm buried silicon dioxide (SiO2) layer is patterned and etched to produce the waveguides and resonators; after additional deposition of SiO2 and planarization, 300 nm of nichrome is deposited and patterned to form the heaters. Finally, another layer of SiO2 is deposited to protect the heaters from outside elements. The waveguides, which are 450 nm wide and 250 nm thick, are adiabatically widened and narrowed at all intersections, which reduces crosstalk (simulations show a crosstalk of -20 dB).
The heaters can tune the microring resonances over a 10 nm wavelength range, which is larger than the rings’ 8.8 nm free spectral range. The resonances have a full-width half-maximum bandwidth of 0.31 nm (38.5 GHz). Tests on the prototype router showed a maximum extinction ratio of 20.79 dB for a routed signal.
“We haven’t tested the speed of the heaters,” says Nicolás Sherwood-Droz, a Cornell researcher. “However, it’s been shown many times that thermo-optic switching at these small feature sizes can work below a few microseconds—plenty of speed for these types of networks. Note that the network here is to be used as a ‘circuit-switched’ network, which doesn’t need to switch as often as a packet-switched network. The emphasis is on passing large amounts of data at once, as opposed to in packet form.”
Another switching technique relying on the electro-optic (EO) effect, which has a much faster switching speed of 100 ps, could be used instead of the thermo-optic effect. But the tuning range using the EO effect would be only about 2 nm, which is less than the rings’ free spectral range; as a result, shifts of greater than 2 nm in the rings’ resonant wavelength due to fabrication-related misalignments of the rings would render the rings unable to tune to the source wavelength.
In a practical on-chip network containing these routers, the total thermal power produced by the heaters would be inconsequential when compared to the thermal power produced by the processors on the chip, notes Sherwood-Droz. The overall chip temperature would likely be maintained at a constant by a thermoelectric cooler attached to the chip.
The 4 × 4 router “is effectively a ‘tile’ that can be repeated as part of a larger network,” says Sherwood-Droz. “The benefit of such a network is that it is in fact scalable. To expand, one would add more of these routers, along with gateways in between (to input and export data from the network), rather than designing a completely new device.”
- N. Sherwood-Droz et al., Optics Express, 15915 (Sept. 29, 2008).