When Electrons and Photons Converge

We recently had the opportunity to catch up with David Miller, one of the industry's optical interconnect "gurus." Miller is currently the W.M. Keck Foundation Professor of Electrical Engineering at Stanford University, and is director of the E. L. Ginzton Laboratory and the Solid State and Photonics Laboratory at Stanford. Previously, he worked at AT&T Bell Laboratories in the Advanced Photonics Research Department.

His research interests include physics and applications of quantum well optics and optoelectronics, use of optics in switching, interconnection, computing, and sensing systems, and fundamental features and limits for optics in communications and information processing.

SI/EP&P: ICs have traditionally used interconnects made of aluminum and silicon dioxide and are moving to copper and low-k. What is your perspective on how optical interconnects might fit into that scheme in the near future?

Miller: Obviously, there's a question of where in the system are you going to use optics. Are you going to put optics right down onto the silicon chips? Are you going to do connections on the chips with optics, or are you going to do connections off the chips with optics? Are you willing to do connections off the boards? To do connections off the boards, the backplanes, that's a technology today. You can use optical interconnects for that now. For example, it's used quite a lot in high-performance switching systems, which tend to be systems that just can't avoid sending very large amounts of data. One of the things that drives them to do that is to get enough density in connections on the boards ... especially as you start going long distances.

1. David Miller in his office at Stanford University, talking about the future of optical interconnects.

It's not going to change the silicon-silicon dioxide transistors. It's not going to change the aluminum to copper or even the copper/low-k. But if you start looking at doing connections off-chip, especially in difficult jobs like switching systems, then you start thinking very hard about going to optics. My personal opinion is, when it gets down to the chips, certainly for getting information off the chips, it will certainly be a hybrid technology of some kind, probably a solder bonding technology.

SI/EP&P: Does that mean you would be generating the light on the chip?

2. An experimental setup in the Gintzon Lab at Stanford University, where light beams are sent from one silicon chip to another, achieving full logic level to logic level connectivity.

SI/EP&P: What about optical devices, like DWDMs and micromirrors and such. Do you see a trend in integrating more of those optical devices into a single planar device?

Miller: Making optical devices using silicon platforms is how some people are addressing the emerging generations of components. Several companies do that — things like arrayed waveguide gratings are made that way. To the silicon person, these things are gigantic. They occupy a lot of silicon real estate, several centimeters on a side. The optics people like that because it's a lot smaller than what they had before. To the silicon people it's absolutely gigantic. Yes, there is going to be a trend toward integrating more function onto those things. I think what's going to happen first is people will try to integrate rather straightforward things; they're trying to integrate things like controllable attenuators in with the wavelength splitters, or trying to integrate things like monitoring photodetectors so you can tell how much power is in each channel. That will be happening.

One of the challenges in optics is the whole issue of packaging; packaging the optics, packaging the fibers onto those things. That remains as one of the major cost issues in the whole optics business.

SI/EP&P: Is that mostly because of the difficulty in lining things up?

Miller: Yes. What you're trying to do basically is align a cylinder to the edge of a plane within a fraction of a micron. That's possible, but it's not easy. It may be very good that the semiconductor manufacturers are coming into this business since it means that organizations that do a lot of manufacturing are trying to figure out how to solve some of these manufacturing problems. I think that will be very important. I think it will happen, and as we move out into the network — as you move to the optics coming down to shorter distances — by that I mean replacing your TV cable to your home, or replacing the phone line to your home. Those are very cost-sensitive areas, and the issue of getting the cost down in the manufacturing of optical components will be very important. That obviously drives integration and it drives good manufacturing of the packaging.

But I think it will be a while before you start seeing a lot of electronic functionality built onto those chips, although eventually that will come. That's not the major problem in the short term.

SI/EP&P: In terms of adding logic circuitry?

Miller: Yes, that's not the problem for the medium term. Eventually, they'll add that in. And certainly from a university point of view, that is the kind of thing we think about quite a lot. We think about how eventually are we going to bring network photons right down onto silicon chips. In the long term, that's a very exciting prospect because there's the potential for collapsing whole racks of equipment onto one chip. I think that it can be done; it's a technical challenge. It's something we have to think a lot about and figure out individual ways to deal with that.

SI/EP&P: These are mostly passive components now, but you were talking about controllable attenuators. Do you see in the future that you'll be using electrical signals to alter the optical properties of the material and doing switching that way?

Miller: Already, people do some switching that way. There are optical switches that are very good switches where you put a voltage on them and they will send a signal from one waveguide into another instead. If you're working on a silicon-based technology, there aren't any good fast ways of doing that, but some of the switching doesn't have to be fast. Some of the switching can be in the 10 msec timescale. There's also the possibility of doing some of the switching with micromechanical switches. There's lots of different levels of switching in the system. Provisioning and protection switching doesn't need to be fast, but once you start talking about packet switching, you're down into the 10 nsec level. Then the idea of being able to integrate electronics and optics rather closely so that you can do the switching electronically but get the information in and out optically, that becomes very attractive in the medium to long term.

SI/EP&P: You were talking about bringing network photons down onto the silicon chip. Is there a way that you envision doing that with silicon or do you have to go to III-Vs at some point?

Miller: For the moment, we have to go to III-Vs. That's not necessarily bad. I think in general when you integrate things — really, really integrate them — then you have to compromise everything. So if I were to insist tomorrow that I was going to put optical devices on a chip, even if they were silicon optical devices, I'd have to compromise my CMOS. The idea of being able to do a hybrid integration is a very attractive one for a number of reasons. You can get better performance out of it.

There are several problems with silicon as an optical material. Basically, other than photodetectors in the visible or very near infrared, it's not a good optical material. If you make photodetectors in CMOS, for example, they are actually not very good photodetectors at any wavelength. They will work, but they are useless at telecommunication wavelengths. That's one problem with using silicon to detect light: It will work but not very well in an IC process at short wavelengths; it won't work at all at telecommunications wavelengths.

Then when you start talking about either modulating or emitting light, silicon is not really good at it at any wavelength. It would take a pretty fundamental breakthrough to change that. There are quite a lot of people that do good research work looking at getting light out of silicon, and there are many interesting ideas that people have been playing with. That's a good thing to look at. But if you're talking about getting high-bandwidth information out in optical form, you're either going to get it out by emitting the light or by modulating the light. Getting silicon to emit light is hard but not quite impossible. But the light that it emits is incoherent light. It's like a light-emitting diode, not a laser. From a practical point of view, if you start talking about anything like optical interconnects, if you want to do them fast, and you want to do any reasonable density of them, I don't think you can do it with LEDs. Light-emitting diodes just won't work. They are optically too inefficient and you can't modulate them fast enough.

Silicon (research) is at the stage of trying to make something that's an interesting LED. And there could be many good reasons for doing that — it's great stuff to do in research. But even if you made a good LED, it would not solve the problem of getting light out of silicon for dense interconnects. For that, you need to make a laser, which is much harder — or some kind of modulator. And again there is not a good fast modulator in silicon.

SI/EP&P : Last year, Motorola announced it had solved the GaAs-on-silicon problem with an intermediate layer. Do you see potential with that?

Miller: I don't consider myself an expert on that, but even if they have solved that problem, it doesn't mean that's how you're going to do optics. At Bell Labs when I worked there, a guy that worked for me there by the name of Jack Cunningham was actually quite successful in doing GaAs on silicon. We were able to make modulators and photodetectors in a GaAs-based system right on silicon. It worked perfectly well. That was about 10 years ago. Then we realized that this wasn't going to solve our problem for doing something like optical interconnects on and off the chip, because there are too many process incompatibility issues. If I tried to grow GaAs on an actual silicon CMOS run, I'd have an awful lot of trouble doing that. You'd have to grow the GaAs structures at an intermediate stage in the silicon CMOS fabrication, and then you'd have to put that stuff back into the silicon line. The process compatibility issues in doing that are pretty substantial.

SI/EP&P : What kind of work are you doing here at Stanford?

Miller: We do optical interconnects right in and out of silicon chips. We are actually able to send light beams from one silicon chip to another and do full logic level to logic level connectivity. We do that by solder bonding arrays of III-V devices onto the surface of silicon chips. We put these on pretty densely — over the entire surface of the chip, they are only separated, center-to-center, by 62.5 µm. And that's not the limit. So we do dense optical interconnects right into silicon chips to try to understand what happens if you try to do that — what are the problems, what are the benefits. We get it to work pretty well in the lab. I think it is going to be a progressive and interesting technology as we start pushing the clock rates well up into the gigahertz range. Electrical wiring problems become severe and the idea of going to optics is very attractive.

We also work on various other things, many of them related to telecommunications. For example, we are looking at a wavelength converter, which is the kind of thing we'll need wavelength division multiplexed optical systems. Because at some point in the networks, you'll need to be able to change the signal from being on one color to being on another, just so you can work the network efficiently. That device actually works; it's like an optoelectronic logic gate. It puts a light beam in that generates a little electrical signal locally inside the device, and that little electrical signal is used to drive a modulator, all inside the one structure. So we can get that phenomenon to occur in picosecond timescales. Currently, we're looking to see if that will make a practical device at the 10 Gb/sec level.

SI/EP&P : Do you work with nanotechnology?

Miller: There are other completely different kinds of things we do, looking at sensing and digital-to-analog conversion. But we've been very stimulated by some of these nanotechnology ideas and so we've been looking at ways of using photonic crystal ideas to make wavelength splitters. We have a nice trick that we can do there, is that we can use existing fabrication methods in the optics industry to make our nanostructures. Nanotechnology is getting increasingly interesting, but I think it has a way to go if we try to think of something like photonic crystals being a platform for doing everything we want to do. That's still some way off. But it's a very interesting concept for miniaturizing components or other ways of designing components. It's going to be a very interesting field.

SI/EP&P: Do you see any hope for optical computing?

Miller: I've worked in optical computing for a long time and I came to the conclusion after a while of doing that that we were always making a mistake by trying to compete with silicon, and that we were doing the right thing when we were trying to compete with copper. So much of my own work in that field has leaned toward optical interconnects. The physical reasons for wanting to do optics in interconnects are very strong; it's a matter of getting the technology soon enough at that price you want. But if all things were equal, you'd use optics for interconnects over any distance.

I do think it's interesting to look at small numbers of very fast devices, operating in a time/speed range where it's difficult for silicon to operate, and operating so that you're making a very good conversion of the optics into logic and then back into the optics again. Wavelength converters are examples of that. That's sort of an optical computing device. It's not trying to do a very complicated job. It's trying to do a fairly simple job, but do it very, very fast and doing it pretty well directly on the light beams. I don't see us having an optical computer that's going to replace a silicon computer. That's not going to happen.