Get email delivery of the Cadence blog featured here
Last Tuesday was the first day of the virtual event CadenceCONNECT: Photonics Contribution to High-Performance Computing. Note that the subtext of this is actually using silicon photonics for performing high-performance computing, not just providing photonic communication links within data centers.
The opening keynote was by Odile Liboiron-Ladouceur. She did her undergraduate at McGill University in Montréal, Canada, and then and her MS and PhD at Columbia University in New York City. She worked at Teradyne in Boston from 1999 to 2000 and then joined Texas Instruments in Dallas and spent two years working for its fiber-optic business. She was then a summer internet at "Watson", the IBM T.J. Watson Research Center. She then returned to McGill where she is currently an Associate Professor and Canada Research Chair in Photonics Interconnect. Ironically, given that job title, her presentation was titled Photonic Integration – From Switching to Computing.
Her presentation started being more of a tutorial on photonics integration, before moving onto "the failed promise of optical switches" and wrapped up with "the promising optical processor".
The above pictures sums up the promise of optical systems on chip (SoCs). On the left is a 2008 design that is 2m by 1m. On the right is a 2015 design with a footprint of 0.02m2.
The attraction of CMOS integration is that it can leverage the infrastructure for high-volume silicon manufacturing, which makes it more economically viable. The manufacturing process for SiPh is "more or less" silicon compatible.
There are several SiPh foundries with MPW shuttles. However, there remain several manufacturing challenges:
There are various pros and cons to silicon photonics but the big one is the lack of a laser. One big advantage of optics of electrical is that they can be multi-frequency ("color") and multi-mode, so several signals can be put down the same waveguide and recovered at the other end.
But the idea was promising enough for the dream of replacing power-hungry electrical switches with optical switches: in the data center, the top-of-rack switches, the mysteriously titled ROADM (reconfigurable optical add-drop multiplexer), or perhaps even on motherboards. This is where Odile said it was a "failed promise". There has been work going on at the motherboard level but electronics always finds a way to go faster (like the Mellanox chip in the lower left, which she described as "our competitor").
The Mach Zender Interferometer (MZI) was invented by two physicists to enable the phase measurement of a coherent light source. You might think that Mach and Zender are a couple of guys in a modern lab at MIT, but this is an invention from 1891. You can still read the original papers online. The basic idea is that you split the light source in two, the two components travel on different paths, and you recombine them.
It turns out the MZI can be integrated. Depending on the difference in the two light paths, the waveforms either go straight through ("bar") or swap ("crossbar"). They can then be scaled to multi-stage topologies.
So how do you control an MZI? The main way is by temperature. Heat up the waveguide by metal or by doping the waveguide and then using resistivity (put a current through it as well as light).
By combining multiple-stage, you can build an optical switch like the above 4-input 4-output switch, controlled by six voltages. However, in practice, optical switches remain challenging. The elephant in the room is that you still need electronics to control the switches, which raises the question of how useful it is to convert signals to optical for switching. Process variation is also harder than with electronics, reducing the attractiveness.
However, beyond interconnect, photons can be used to compute. Optical processors have been the buzz since 2015. MZI performs a linear transformation of the input signal, in effect multiplies it by a 2x2 matrix. As with the switches (or computer graphics for that matter), transformations can be stacked on each other and all the matrices are multiplied together.
But all is not rosy:
Photonic computing can do WDM (wavelength division multiplexing) and in-memory compute in a way that avoids the von Neuman bottleneck. One application is high-frequency trading, which if you've read Michael Lewis's book Flash Boys, you will know is very valuable.
You might have noticed that doing lots of matrix multiplication sounds quite like neural network interference, and that is one of the attractive areas. Photonic computing has the promise to deliver 25 times higher tera-ops per watt than the current solution of an NVIDIA GPU.
Photonic integration contributes to greater capacity in communications, today mainly in transceivers but maybe one day in switches.
Photonic computation is an open field of possibilities for research and startups are piling in.
And finally, here's her team from last year:
Sign up for Sunday Brunch, the weekly Breakfast Bytes email.