Nearly all long-distance communication now depends on optics, and all-optical technology is gaining a larger market share in short-range networks. Future local-area-network technology such as 10 Gigabit Ethernet will make very heavy use of optics. Meanwhile, optical data links inside large electronic systems connect frames of equipment to peripheral devices, such as storage area networks.
Despite these advances, it still would seem a substantial leap to propose that we might use optical connection all the way down to silicon chips themselves. Current optical connections usually span at least tens of meters and are often much longer. Why would we consider making optical interconnects to the chip, where connection lengths might be as short as centimeters or even millimeters? If we do think it is a good idea, how would we make such interconnects?
First we need to understand some of the problems of electrical wiring, as well as the potential benefits of optics.1 One key issue is the scaling of the number of bits per second that an electrical connection can support. For a simple line limited by its resistance R and capacitance C, we can send about one bit of data for every time period equal to the product RC. This behavior governs most of the electrical connections on silicon chips. In general, if a given wire cannot handle the data rate we need, we can make the wire thicker, reducing its resistance. It's an easy fix, but once we have filled all available space with wiring, we can no longer take that approach.
For high-speed digital systems, all signal lines behave as inductive-capacitive (LC) lines, except for lines on chips. The longer lines on chips are now also unavoidably starting to change toward LC lines. Surprisingly, LC lines, such as coaxial lines or other impedance-matched lines used for high-speed signals, obey the same kind of scaling relation as RC lines, for reasons also related to their resistance. Even more surprisingly, for simple on-off signaling, such LC lines actually can carry less information per unit cross-sectional area than can the RC versions, so changing to impedance-matched LC lines actually makes the scaling problem worse.
Once all the available space becomes filled with wiring, suddenly there is a problem at all length scales in the system. Miniaturization cannot solve the problem. Though interconnects on silicon chips are arguably the best electrical interconnects there are, the Semiconductor Industry Association predicts that after 2006 no known physical solution will be able to scale to keep up with the speed and integration level of the chips required at that time.2 This scaling barrier is the underlying reason why even those on-chip interconnects are starting to run into serious limits.
Hence we may need to consider short- distance optical interconnects much sooner than we would have guessed. Optical interconnects are not bound by the electronics scaling limit because the physics of the two approaches are completely different.
Off-chip interconnects have even greater problems than on-chip versions. On the chip, it is possible to break up long interconnects using repeater amplifiers, but that approach is inconvenient off the chip. Off-chip interconnects are thus likely to be an attractive first implementation of optical interconnects to silicon chips.
Optical interconnects are interesting for many other reasons.3 They are immune to electromagnetic interference and crosstalk, provide voltage isolation, and can have very low loss. We also can consider using free-space optics, imaging many light beams from one chip to another, for example. This approach is essentially unthinkable with electrical interconnections, each of which must have its own distinct wire. the promise of CMOS
The idea of optically interconnecting directly to silicon complementary-metal-oxide-semiconductor (CMOS) chips is attractive for several reasons. CMOS will be the dominant technology for commodity electronics in the future. CMOS processing has the potential to dramatically reduce the cost of optical interconnects with respect to expensive custom electronic approaches. The integration of optoelectronic devices with CMOS also can improve the performance of the interconnect system. Integrating photodetectors using CMOS chips can reduce their capacitance in comparison to other approaches such as wire bonding. Reduced capacitance yields receiver circuit designs with lower power, higher-input voltage swings, and greater noise immunity.
Historically, a great impediment to integrating optical interconnects with CMOS was the absence of suitable optical output devices. Fortunately, there are two devices that may be able to fill this rolequantum-well-modulator diodes and vertical-cavity surface-emitting lasers (VCSELs). Of the two, quantum-well diodes have been integrated in larger arrays and have been used more frequently at a systems level. Their disadvantage is that they need an external laser source, but even this offers advantages in that the external source can have a well-controlled wavelength and also can be set to clock the interconnect itself.
Figure 1. An array of 200 quantum-well modulator diodes, measuring 62.5 X 125 µm, bonded onto a CMOS chip, can modulate light beams perpendicular to the chip.
Both the VCSEL and the quantum-well modulator are made with III-V materials such as gallium-arsenide and aluminum-gallium-arsenide. As such, they require some integration technology to link them to the silicon circuits. The approach we have taken has been to solder-bond arrays of quantum-well diodes to finished silicon CMOS chips, an approach pioneered at Bell Laboratories (Murray Hill, NJ).4 Intended to work with arrays of light beams perpendicular to the surface of the silicon chip, these devices can function as both detectors and reflection modulators (see figure 1). Using this technique, we have interconnected one CMOS chip to another with light beams reflected off modulators on one chip and received on photodiodes on another.
In one example, the modulators are read out using short optical pulses from a clocked laser system (see figure 2).5 The drive signal has been degraded with jitter, often encountered in electrical systems. Reading the modulator with a short-pulse laser in the middle of the modulation window allows the signal to be transmitted while the jitter is removed, an unusual and special feature of such an optical interconnect.
Figure 2. A modulator drive signal (top) has jitter added to simulate the effect of electrical systems, but the short-pulse optical source used in receiving and optically retransmitting the signal removes the distortion (bottom).
Another example uses wavelength division multiplexing to make a multichannel interconnect between chips over a single fiber with only one laser (see figure 3). A short-pulse laser generates a broad band of wavelengths. The reflective quantum-well modulators on the transmitter chip then individually modulate separate bands of wavelengths. The grating combines the wavelengths so they can be sent over one fiber to similar optics that in turn separate out the wavelengths so the receiver chip can read them out individually.6 These are just a few examples of the many research projects underway on optical interconnects.7
Figure 3. Reflective quantum-well modulators on the transmitter chip modulate separate wavebands, which are combined for transmission and demultiplexed on the receiving end.what the future holds
Optical interconnects to silicon CMOS are still just a research topic, and it's difficult to say when they will be implemented. But the problems with interconnects in high-performance electrical systems will become severe later in this decade, and some substantial new approach to interconnects will be needed. Optics is arguably the only physical solution to such problems.
Additionally, in about the same time frame, silicon CMOS interconnects will be running at clock rates that will make them fast enough to interface directly with optical channel bandwidths, for example, at 10 GHz. Spectral bands 12.5 GHz wide soon may be practically available for filling the entire optical fiber bandwidth. This creates the exciting prospect of the convergence of silicon CMOS and optical networkingCMOS optical interconnects could become optical networks. Much work would remain to make the necessary technology a practical reality, but the payoff of that convergence could be substantial for both the electronics and photonics worlds. oeReferences
1. D. A. B. Miller, Proc. IEEE 88, 728-749 (2000).
2. "International Technology Roadmap for Semiconductors, 1999Interconnect," Semiconductor Industry Association, 1999.
3. D. A. B. Miller, Int. J. Optoelectronics 11, 155-168 (1997).
4. A. Krishnamoorthy and K. Goossen, IEEE J. Sel. Top. Quantum Electron. 4, 899-912 (1998).
5. G. Keeler et al., IEEE Photonics Technol. Lett. 12, 714 -716 (2000)
6. D. Agarwal, G. Keeler, B. Nelson, and D. A. B. Miller, IEEE LEOS Annual Meeting, Paper #ThT4, (1999).
7. Y. Li, E. Towe, and M. Haney, eds., Proc. IEEE 88, No. 6, 723-863 (June, 2000).
Taking a quantum leap
David Miller and associates at Bell Laboratories (Murray Hill, NJ) performed seminal research into quantum well modulators during the early 1980s. His group discovered the quantum-confined Stark effect, which is a way of changing the amount of optical absorption one can achieve from quantum materials when applying voltage. "These devices are now extensively used in telecommunications for externally modulated lasers," says Miller.
The group also led the first wave of investigations into optical interconnects. Miller personally experimented with the physics behind optical interconnects, trying to convince himself that they were viable, and found that optics act as an impedance transformer. By nature, all electrical transmission lines are low impedance, but electronic devices are high impedance, which creates a fundamental mismatch. "If you don't match impedances, you get lots of reflection and wasted power," Miller says. In the future, optics may be used to save power within interconnects, a process known as quantum impedance conversion.
Originally, there was a good deal of justified skepticism about optical interconnects, says Miller, especially since there was no need for them. "But physics told us that eventually there would be a scaling problem with electrical interconnects. It just hadn't hit yet. Now it is starting to bite." It is possible to make faster transistors on a chip, but the underlying problem is that interconnects can't keep up. "It is like making cheaper, faster cars but you can't go any faster on the roads," he continues. "Optics may not be the ultimate answer to this, but it is the only physical solution that I know of. Without optics, you may have to redesign the entire system."
In 1996 Miller left Bell Labs to teach at Stanford University (Palo Alto, CA). "I like being a professor a lot, but it is actually harder than working at a company," Miller says. He is the W. M. Keck Foundation Professor of Electrical Engineering at Stanford, as well as the director of the E. L. Ginzton Laboratory and the Solid State and Photonics Laboratory at Stanford.
David A. B. Miller is director of the Ginzton Laboratory at Stanford University, Stanford, CA.