At DesignCon at the end of January, there was a panel on photonics. The title was Photonics Coming of Age: The Emergence of PDKs. The panelists were:

• Moderator James Pond of Lumerical (CTO). Lumerical provides photonic simulation software (and more) and is a partner with Cadence for an integrated silicon photonics design flow.
• Gilles Lamant of Cadence (leading our work in photonics). I'm not going to introduce Cadence.
• Mohamed Youssef of Mentor/Siemens (foundry enablement). Mentor, now part of Siemens, is a leader in physical verification, an important part of a silicon photonics flow with its odd geometry with lots of curves.
• Samir Choudhry of TowerJazz (design enablement). TowerJazz is a foundry.
• Ashkan Sayedi of HPE Labs (working on PDKs for photonics). HPE is HP Enterprise, computing for companies (as opposed to HP, which retained the consumer stuff like PCs and printers).
• Rui Santos of SMART Photonics (principal engineer). SMART is a pure-play InP foundry.

Introduction to Photonics

Since not everyone in the audience was really up to speed on silicon photonics, James started with an introduction. Photonics, he said, is:

everything to do with light and its manipulation.

The first basic building block is a waveguide, that lets you guide light around the chip. This can be made from silicon and also from indium phosphide (InP). It can support multiple modes (polarization, frequency). Light doesn't like sharp corners, but you can guide it around a curve (5um in silicon) with very little loss. Once you have a waveguide, you can couple and split, a bit like optical plumbing.

The active building blocks (waveguides are passive) allow the phase of the light to be controlled electrically, either with a PN junction right in the middle of the waveguide, or with thermal phase control using a heater. This allows electrical signals to be converted to optical signals. In the other direction, photodiodes are used to detect light, so turn optical back into electrical. This usually requires a germanium layer to be deposited on top of silicon. There are two ways of handling the light source. One is to do that off-chip (or on a separate die in the same package). Alternatively, with InP it can be done on-chip naturally, but with Si it is harder and generally involves attaching III-V material.

For a more detailed introduction to photonics, see my post about the recent Cadence summit Diwali, the Hindu Festival of Lights...and Photonics, the Silicon Festival of Light and other earlier posts linked from there.

For manufacture, there is a big tradeoff. Indium phosphide (InP) is a premium material, and gives you a nice integrated laser, which silicon cannot do. On the other hand, if you want 8" and 12" wafers with high-volume manufacturing, then starting from silicon is the best. There are experimental options with InP being integrated into Si. But, in for now:

Si doesn’t have a laser. It is the waveguide router.

The challenge, and the topic of the panel, is that to make something useful today, you need a PhD in photonics. Obviously, that is not scalable, and for every photonic expert, there are a thousand mixed-signal designers. What needs to be done is to bake in a lot of the institutional knowledge and best practices into a PDK (process design kit) to make it scalable, and enable at least basic photonics to be done by the non-initiated. PDKs for semiconductor processes do the same, especially in the FinFET era where it's only a slight exaggeration to say nobody understands the details of the layout rules.

Panel Discussion

Question: What are the fundamental differences between electronics and photonics that will show up in PDK differences?

Gilles (Cadence): There are a lot of common building blocks such as simulation models, both electrical and optical. Both have schematic symbols. Layout generators. All mixed-signal designers know this stuff. One big difference is how layout is drawn, it's all curvy. That is the most visible difference, but there are also a lot of differences due to the fact that signals span a wide range of frequencies (light, electrical, thermal) and cover mulitple modes (electric, magnetic). This brings complexity to the PDK (and the tools). We try to hide it and make the waveguide look like wire, but behind the curtain, it is very different—a waveguide is a device with a model, not just a few Rs and Cs..

Mohamed (Mentor): Electronic PDKs have been around for ages, but are not good at handling variations such as when resistors can go on different layers. then you end up with 300 components. On the photonics side there are fewer variations (no layer variations, no voltages). Electronics is very mature and every device is well parameterized by the foundry and simulated over a range of instances. But on the photonics side, some foundries are still providing black boxes, not parameterized across the full spectrum of parameter values. Physical verification is still needed for photonics, even for curved objects.

Question: What about custom devices that are not in the PDK? Is that just that PDKs are immature? Do we foresee a day when all the devices are in the PDK?

Gilles (Cadence): The components that you see in a photonics PDK are not R, C, Xtor level, but rather couplers, modulators, etc., and so the foundry can't always have exactly what you want. It's as if the foundries gave you a standard op amp, when there would always be something that you'd want to tune. Most of the optical components are at that level of complexity. What foundries can do is give a set of examples, but you're never going to find everything you need without any tuning.

Ashkan (HPE): For electrical, specs and demands get generated in industry, foundry hears that, and they tune the process and the PDK. People who use the electronic PDKs don't know the deep physics—you just use the parameters that are exposed to you. In some ways, the difference between electronics and photonics is just a matter of wavelength. For a transmission line in electronics, which goes a long distance, you know automatically to widen the copper. Analogous things exist in photonics.

Gilles (Cadence): With Spectre-AMS and Lumerical INTERCONNECT, we can not only do mixed-signal but also mix-discipline co-simulation, and we can simulate systems including DSP core, SerDes analog, and optical, all together. Lumerical has used Verilog-DPI to tightly couple the optical simulator to the analog simulator (Spectre) including the true electrical loading of electro-optical devices on the electrical circuitry. So I would not say co-simulation does not exist.

Samir (TowerJazz): When I started to do it myself, it was tough and I ended up putting together a reference flow that was very useful to our customers.

Question: What about test and verification?

Rui (Smart): Verification was a big challenge, too, with equation-based DRC. On the electronics side, there are very few DRCs, and a waiver list that has been pre-negotiated. Tapeout is done in 24 hours. But photonics get 40,000 DRCs. You cannot submit GDS files with 40K DRC errors. The big need for a PDK is that a lot of guys would just draw shapes in MATLAB and not use components that are correct by construction.

Ashkan (HPE): Doing my PhD I was in the cleanroom looking through an electron microscope building it myself. But that’s not going to scale. A lot in electronics reflects 50 years of institutional knowledge. At 7nm and 5nm, design rules are so limited, and we are not going to be there in the photonics world with curved structures. OPC is another whole can of worms for photonics, too.

Question: A lightning round, getting each person's five-year forecast for photonics.

Gilles (Cadence): Something more automated has to be done for things like a 50x50 switch array. Placement automation might never work, it doesn’t even work for analog, but maybe automation will work for routing. As to optimizing devices, very little has been done for robustness to process varation. There is lots of interesting stuff that could be extended to more structures than you see.

Ashkan (HPE): In five years, we will see more photonics in the datacenter. What is on top-of-rack will come down. Pluggable at 5W is too much power,  so it will go on the board. Design tools will mature. Foundries will be able to deliver known good wafers. There will be more schematic-driven co-simulation. We're at an inflection point right now. It’s the technology of the future all the time…but going from 12.5G to 25G datarates, customers demand it. We won’t be able to build electrical 200G SerDes. Photonics is needed to enable it.

Samir (TowerJazz): James showed a graph of the market growing (see above). I hope it is growing faster. There will be standardization. We are the first open foundry (biosense, datacomm, lidar…). Designers will find a way to use “good-enough” processes.

Rui (Smart): We will see a jump in maturity during five-year time. We will see price/cost down from prohibitive cost limited to telecom to devices for datacom, where InP come into its own. Sensing is where the big jump will happen. Customers populating airplanes with photonics sensing. Automotive. Volume markets.

Mohammed (Mentor). One thing we’ll see in less than five years these little small research foundries will need to consolidate and mainstream foundries like TowerJazz and Smart and big foundries will come in. I think we'll see maturity of PDKs in less than five years.

And with that James thanked the panel and released everyone for beer in the exhibit hall.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.