Optical Digital Computers
AT&T Bell Laboratories
Crawfords Comer Road, 46-514
Holrndel, New Jersey 07733
PHYSICAL LIMITS OF SUPERCOMPUTING WHAT ARE WE TRYING TO BUILD?
The fastest switching time for a transistor is around 5 pica The architecture which we are using is based on a simple pipeline.
seconds and yet the fastest computer cycle time is around 5 nano This architecture avoids the Von Neumann bottleneck and is
seconds. Where does this three orders of magnitude in nicely matched with the parallel connectivity and constant latency
performance disappear to? This problem can be traced to the of optics. Such an optical pipeline is shown in Figure 1. It
limited bandwidth and connectivity of electronics. Bandwidth consist of arrays of optical logic gates interconnected via optics.
restrictions limit the speed and add to the design complexity.
Constrained connectivity forces a time multiplexing of the
interconnections which in turn imposes a sequential, address-
oriented transfer of information at the architectural level (Von
Neumann bottleneck), bus level, and even at the memory chip
HOW CAN OPTICS HELP?
Optics has a bandwidth in the Terahertz regime which is
essentially limitless from the perspective of electronics. The
parallelism of optics can also provide an amazing amount of
connectivity. Optics can easily convey an image consisting of a
100 by 100 array of spots. This could be viewed as 10,000
Optics has some other unique advantages.Two light beams can go Figure 1 An optical pipeline consisting of optical logic gates
right through each other without interference. This greatly interconnected via optics.
simplifies crosstalk and EMI problems. One of the more unusual
properties of optics is that traditional imaging systems are
constant latency. All the light from the input image plane reaches
the output plane at the same time. This could greatly simplify
the clock skew problem.
Historically, one of the impediments to the development of an
optical digital computer has been the belief that optical logic
would require far more energy. This has since been disproven PI.
Ironically, it has also been shown that it takes more energy to
communicate with electronics than with optics for distances
greater than 200 microns f31. Smce computation consists of both
logic and communication it would seem from a theoretical point
of view that electronics is now at a disadvantage.
Permission to copy without fee all or part of this material is granted provided
that the copies are not made or distributed for direct commercial advantage,
the ACM copyright notice and the title of the publication and its date appear.
and notice is given copying is by permission of the Association for
Computing Machinery. To copy otherwise, or to republish. requires a fee
and/or specific permission.
0 1989 ACM 089791-341-g/89/001 l/O446 $1.50 Figure 2 A portion of a chip containing over a million lasers.
We are currently pursuing several types of optical logic gates INTERCONNECTIONS
[4~5*6~71.One approach follows from optical interconnects. It A lens can easily convey a 100 by 100 array of spots which could
involves giving electronic chips optical input and output pads. represent 10.000 independent channels. It is easy to seehow this
One approach for achieving an optical output pad involves the use can be used to achieve one-to-one connectivity. If is aIs easy to
of microlasers. We have recently been able to fabricate millions seehow via a beam splitter one-to-two or two-to-one connectivity
of these lasers 181. See Figure 2. We are also working on a can be accomplished. It is more difftcult to see how to achieve a
CMOS optical input pad. random connectivity. The approach which we use relies on a
We are also working on an optical logic gate based on an multi-stage network approach. One approach which we use is
integrated electro-optical &vice called the SEED (self-electrooptic based on a perfect shuffle [lo]. A more optically efficient
effect device) 141. The device is functionally equivalent to a approach is based on crossover networks [ 111. Such a network is
latching NOR gate. Arrays of such devices are shown in Figure shown in Figure 4. Either of these approaches is capable of
3. The SEED can also be used as a modulator and thus used as an supplying thousands of very high bandwidth, low energy, constant
optical output pad for electronics. latency interconnections.
Our original attempt at building one stage of an optical pipeline
took three, 4 by 12 foot optical benches. Our current approach
takes about 1 square foot [121 and is shown in Figure 5. We are
now investigating using integrated free spaceoptics to reduce this
setup to several squareinches 1131.
Figure 3 An array of SEED optical logic gates.
Figure 5 Our current optical pipeline stage.
We have developed several techniques for converting circuits into a
pipelined, multi-stage network. Given any circuit we can show
that it can be representedwith at most one extra level of logic and
w i/t ” Mirror approximately one third greater width [141. Such a circuit is
shown in Figure 6.
vj Mask for prism array
- Optical logic array
Figure 4 An optical crossover network.
Such a circuit as shown in Figure 7 cart be “regularized”, cast into
a regular array as shown in Figure 8.
Figure 6 The use of a crossover network to route a
combinatoric circuit. Figure 8 A regularized combinatoric circuit
This circuit can then be folded down with the aid of deIay lines
into the circuit shown in Figure 9. Computational origami can
The question occurs as to how to represent circuits which are be used to fold systems as well as jusr simple circuits.
wider or longer than the pipeline. A technique called
computational origami has been developed which involves the
reformatting of computations [15J. It takes a computation,
regularizes it, and then folds it into a format which is more
suitable for processing. .
r lik.- .I
Figure 9 A folded version of the combinatoric circuit.
Figure 7 A typical combinatoric circuit.
CONCLUSION [ 1l] A. Dickinson and M. Prise, “Crossover Networks and Their
In terms of raw speed, it is believed that the output of an Optical Implementation,” AppIied Optics. ~0127, pp 3 155-3160,
electronic chip will be limited to around a Gigibit. The use of 1988.
optical output pads should be able to extend this limit to around
10 Gigibits at which time the speed will be limited by the wires  “A Module for Optical Logic Circuits Using Symmetric Self
on the integrated circuit itself. By modifying the architecture and Electrooptic Effect Devices,” M. Prise, R. LaMarche, N. Craft,
giving each logic gate art optical input and output capability this M. Downs, S. Walker, L.Chirovsky, and A. D’Asaro. submitted
limit should be able to be pushed to around 100 Gigibits at which Applied Optics, July 1989.
time the speed will be limited by some of the intrinsic properties
of the semiconductor. Faster optical non-linearites exist. They [ 143A. Dickinson and M. F’rise, “Planar integration of free-space
are weak but they react in order of femto-seconds (10-15). Their optical components,” Applied Optics, vol 28, pp 1602-1605,
use is speculative at present but they are being studied. 1989.
In terms of parallelism, it is believed that optics can easily  M. Murdocca and T. Cloonan, “Optical Design of a Digital
achieve over 50 times more connectivity. This should open up Switch,” Applied Optics, vol. 28, no. 13, pp 2505-2517, July
some of the architectural bottlenecks. Ideally, this should have a 1989.
direct effect on throughput.
 A. Huang, “Computation Origami - The Folding of Circuits
and Systems,” Conference Proceeding of the 1989 Optical
Computing Topical Meeting of the Optical Society of America,
REFERENCES Salt Lake City, pp. 132-135. Feb. 1989.
[l] A. Huang. “Architectural Considerations Involved in the
Design of an Optical Digital Computer,” Proceedings of the
IEEE, vol. 72. no. 7, pp. 780-786, July 1984
 D. A. B. Miller, J. E. Henry, A. C. Gossard, and J. H.
English, “Integrated Quantum Well Self-Electroopptic Effect
Devices: 2 x 2 Array of Optically B&able Devices,” Applied
Physics Letters, ~0149, pp 821-823, 1986.
 D. A. B. Miller, “Optics for low-energy communications
inside digital processors: quantum detectors, sources, and
modulators as efficient impedance convertors,” Optics Letters, vol
14(2), pp 146. 1989.
 A. L. Lentine, H. S. Hinton, D. A. B. Miller, J. E. Henry, J.
E. Cunningham, and L. M. F. Chirovsky, “Symmetric Self-
Electra-Optic Effect Device: Optical Set-Reset Latch,” Applied
Physics Letters, ~0152. pp 1419-1421, 1988.
(51 J. L. Jewell, A. Scherer, S. L. McCall, A. C. Gossard, and I.
H. English, “G&z-A& Monolithic Microresonator Arrays,”
Applied Physics Letters, ~0151, pp 94-96, 1987.
(61 L. C. West, “Picosecond Integrated Optical Logic,”
Computer, pp 34-46, Dec. 1987.
 M. C. Gabriel, ‘“Transparent nonlinear optical devices,”
Optoelectronic Materials, Devices, Packaging, and Interconnects,
Ted E. Batchman. Richard F. Carson, Robert L. Gallawa, Henry
J. Wojtunik, Editors, Proc. SPIE 836, ~~-222-227 (1987).
 J. L. Jewell, A. Scherer, S. L. McCall, Y.H. Lee, S. J.
Walker, J. P. Harbison, and L. T. Florez, “Low Threshold
Electrically-Pumped Vertical Cavity surface Emitting Micro-
Lasers.” to be published in Electronics Letters.
 A. Dickinson and M. Prise, “A Free Space Optical
Interconnection Scheme,” Conference Proceeding of the 1989
Optical Computing Topical Meeting of the Optical Society of
America, Salt Lake City, pp. 132-135, Feb. 1989.
[IO] A. Dickinson and M. Prise, “Optical Implementation of
Perfect Shuffle Interconnections,” Applied Optics, vol. 27, pp