eugene@wilbur.nas.nasa.gov (Eugene N. Miya) (08/16/90)
Dr. David Kahaner is a numerical analyst visiting Japan for two-years under the aspice of the Office of Naval Research-Far East (ONRFE). The following is the professional opinion of David Kahaner and in no way has the blessing of the US Government or any agency of it. [DKK] Back issue of most reports will shortly be available via anonymous FTP. Host is pending. [ENM] This report is cross-posted to comp.arch, perhaps the most relevant news group. To: Distribution From: David Kahaner ONRFE [kahaner@xroads.cc.u-tokyo.ac.jp] Re: Optical Computing in Japan 7 August 1990 ABSTRACT. Optical computing activities in Japan are surveyed. INTRODUCTION. In recent years the field of optical computing has been rapidly broadened into various areas, such as investigations of optical analog and/or digital data processing, and optical and optoelectronic phenomena and devices for optical computation. An optical computer is a computer in which light is used somewhere. This can mean fiber optical connections between electronic components, free space connections, or one in which light functions as a mechanism for storage of data, logic or arithmetic. The main motivation of many recent studies of optical computing is the increasing interest in developing a new parallel computing system capable of processing large amounts of data at high speed, and my own interests in the subject are centered on this potential application. My aim was to discover how close optical computing is to being of use to the constituency of numerical computing that I represent, and to gain some understanding of the ways digital optical computing and neural computing overlap. In my opinion most scientists engaged in mainstream scientific computation have little knowledge of neural computing and even less of optical computing. Nevertheless, these are well established scientific fields with thousands of researchers, professional societies, journals, and international meetings. For example, Optical Computing '90, held in Kobe Japan 8-12 April 1990 was a major conference on this topic with almost 500 attendees. (See below for a summary and evaluation.) Optical computing is seen by a number of Japanese as an essential direction for computing research. Here are some examples of their comments. MITI: "electronics is the science of the twentieth century, and optics is the science of the twenty-first." Dr. Izuo Hayashi, Director of the Optoelectronics Technology Research Laboratory (OTL) in Tsukuba: "The combination of photons and electrons will create new kinds of system which we cannot imagine just using an extension of today's technology. For instance, imagine real 3D integration, by which I mean wafer-to-wafer communication, vertically, by light, so that we can make stacks of hundreds of wafers by integration. Once we master optoelectronic integration technology we can begin to imagine new architectures." Concerning the research activities in Japan and the U.S: Prof. Takeshi Kamiya, Department of Electronic Engineering, University of Tokyo: "Compared to the U.S. we have a wider variety of research groups in Japan that are developing devices dedicated to optical computing", while "In the U.S., I think there is a wider variety of groups looking for new architectures for optical computing." Dr. Ken-ichi Kitayama, supervisor at NTT Transmission Systems Laboratories: "[ATT Bell Labs] are looking at the short term target rather than the longer term target, and it seems that all their efforts are now concentrated on 1995." "But NTT's long term goal is to establish optical processing technology and to fully exploit massive parallelism by optical means. This will produce new types of optical devices and an optical architecture. We expect that in long run, research in a broad range of areas will be fruitful. So for now, we consider this to be a basic research phase, not the practical development phase." "The photoelectronic or optoelectronic computer is the direction of the future". WHAT IS OPTICAL COMPUTING? The concept of passing light through lenses to perform computations is not new. I took a course on this in the 1960's. The fundamental idea can be illustrated by noting that a simple lens essentially performs a two dimensional Fourier transform of its input in real time for arbitrarily complicated image, whereas using digital computation the effort increases rapidly with the number of data points or pixels. Using a lens in this way is an entirely analog process, and most of the early research considered computation in analog terms very much like it was described in the days of analog computing. In recent years developments have centered on digital calculations, by using optical devices for logic, memory or arithmetic. A stumbling block in this research is that it is necessary to find optical materials that react nonlinearly to input, and thus far sufficiently nonlinear materials have not been available, or their nonlinearities are too weak for practical application. Work is also continuing on using optics to connect traditional circuits. Optical communication has already made a significant impact in computer communication via optical fibers. It is well known that optical fibers have much lower attenuation during transmission than electrical wires in coaxial cables. In addition they are more resistant to electromagnetic radiation along their length. Optical cables are already being used as I/O channels in Japan (Hitachi in 1987, Fujitsu in 1988, and NEC in 1989). Such channels have data transmission rates up to 9MB/second and may be improved to more than double that. In addition they can be used over much longer distances, up to about 1 kilometer for disk channels, about 8 times as far as electrical channels. Prof. J. Goodman (Stanford University), for example, believes that "optical interconnects" in general are promising areas for real products. Further, if these interconnects can be utilized to connect one chip to another (optical output pads) it is speculated that performance in the 10 gigabit range will be possible. Laser beams can cross in arbitrarily complicated ways without losing their individuality, or experiencing crosstalk, at least on larger dimensional scales. Again the Fourier transform provides an excellent example. Each point value of the transform is obtained by integrating over all points in the source plane; that a lens can do this easily is in one sense, the ultimate in parallelism. It is estimated that optics can achieve at least 50 times the parallelism or connectivity of electronic devices. The Japanese expertise in device technology may enable them to capitalize on it better than others. This was aptly summarized by Kitayama in describing research at NTT in optical computing: ``although the applications in the future may be diversified, special purpose hardware may first come in processing images at data rates which are unabtainable using all electronics...One of the promising schemes would be a combination of optical devices and VLSI. Optical neuro-chips may be a longer-term goal...Practical application of optical hardware still seems to stand at the far end of the time line.'' There seem to be four categories of optical computers. (1) Optical analog. These include 2-D Fourier transform or optical correlators, and optical matrix-vector processors. (2) Optoelectronic. These do not yet exist, but would be constructed using optical logic gate devices or 2-D photo diode arrays. The main interest in this type of computing device would be to shorten the pulse delay in chips and other logic elements by using optical interconnections. (3) Optical parallel digital computers. These would use the inherent parallelism of optical devices along with digital electronics for flexibility. (4) Optical neural computer. Specifically designed to implement the massive interconnection requirements of neural networks optically. GENERAL REMARKS. Optical computing is still a branch of experimental optics, with the usual trappings of physical science, i.e., careful attention to fine detail of setup and analysis. The research is not localized in any single country. Early work by scientists at Bell Labs and other US laboratories is now complemented by comparable work in many other countries. As long as optical devices utilize free space the research results are clearly dependent on the planning, creativity, and care of the individual research group rather than on access to technology that is not widely available. To the extent that scientists are beginning to think of optical computing ``chips'' the Japanese researchers have the advantage of access to the substantial resources and basic technological device infrastructure of large Japanese industrial laboratories. The same may apply to work at a very few US labs such as Bell Labs. Thus far, concrete applications are several years away from being useful to the numerical computing community. There seems to be a healthy competition between major researchers. For example Huang at Bell Labs is working on optical logic gates based on a principle he calls SEED, but Kamiya thinks that the NEC approach might have certain advantages such as ability to amplify optical signals. RELATION TO NEURAL COMPUTING. A neural computer, or neural network, is a special kind of highly parallel computer with many computing elements, or nodes performing simple operations (usually just matrix vector product) in a very repetitive manner. In some models of neural computer (which have never been implemented) there are postulated to be tens of thousands of nodes each one of which is connected to all the others. Neural computers compute in the sense that they have streams of input and output bits. They do not require anything resembling ordinary programming; if programming is done at all it is by dynamically changing the degree to which the individual nodes are connected. An important aspect of a neural network is the high degree of parallelism associated with it. Thus it is natural that new parallel computers should seek to implement neural networks as an application (but not the only application). Optical computing researchers believe this parallelism can often be implemented best using optical devices rather than traditional wired circuits. Thus optical computing and neural computing are fields that have developed independently and now sometimes come together for their mutual benefit. Today, most applications of neural computing, and in particular those in which optics play a role are related to image processing, character, target, or voice recognition and similar situations. However several researchers have demonstrated optical devices that can multiply matrices and solve small systems of linear equations, and papers are appearing that attempt to apply neural models to more general reasoning situations. At this time neural networks have not been applied to numerical modeling problems and I have seen nothing on any of my visits to suggest that this is likely. Perhaps these models are fundamentally different from what we usually think of as algorithms. As mentioned above the fundamental operation of a neural computer is multiplication of an input vector (array) into a matrix with elements called the network "weights." Both input array and weight matrix are assumed to have non-negative elements. Let the matrix elements W(i,j) be associated with a two dimensional light mask and the input with a light emitting device array. Let the vector information of the input, v, be radiated as optical intensity from laser diodes or light emitting diodes in such a way that v(j) is radiated uniformly to the j-th column of the matrix W in such a way that W(i,j)v(j) is the light intensity on the back, or output side of the mask at point i,j. This is usually described as "fan out". Then let the light intensity of the i-th row of W be converged onto the i-th component of a light receiving device array, also in a uniform way. By superposition, the i-th component of the output array is then the inner product of the i-th row of W into v, and in this way the matrix vector product is formed. The technical issues to be dealt with include developing appropriate fan out light emitting beams and a mechanism to permit variation of the values of the components of W(i,j). The matrix vector multiplier above is usually part of a "neuron", which takes each output component and returns 1 if it is large enough, 0 otherwise. To do this optically requires optical "thresholding." OPTICAL COMPUTING '90. The 1990 International Topical Meeting on Optical Computing was held 8-12 April, 1990 in Kobe, Japan. Almost 500 scientists attended this meeting. A related meeting on Photonic Switching was held immediately after. Participants came from Japan, USA, USSR, France, UK, FRG, China, Switzerland, Finland, Australia, Bulgaria, Korea, Phillippines, Sweden, Canada, Belgium, and Italy. The meeting was held in the International Conference Center at Kobe, on an island in the Port of Kobe. There were no parallel sessions and the first day was exclusively devoted to three tutorials, by J. Goodman (Stanford), D. Miller (ATT Bell Labs), H. Szu (NRL). A Proceedings, in English, is available from the International Society for Optical Computing by referencing ISBN 4-980813-37-9. I also have a copy of the proceedings. Many of the papers presented at OC90 are variants or extensions of papers that are published in journals such as Optoelectronics or Applied Optics. I am not a physicist and cannot evaluate such aspects of this work. An extensive collection of reprints (in English) was sent to me by Professor Yoshiki Ichioka, Department of Applied Physics, Osaka University, 2-1 Yamadaoka, Suita, Osaka 565, Japan, for which I am very grateful. Ichioka and colleagues have been concentrating on building optical devices that will perform fundamental logical functions, AND, OR, NOR, etc. The group has created several new ways to implement some logic devices. Their computer design using these techniques is called OPALS. Work on this has been going on since at least 1983 and is well known outside of Japan. However, implementation requires development of several new kinds of devices, so this research is several years away from practical application. Goodman and colleagues from Stanford presented a paper on a simulated 64- node shared memory multiprocessing system, the IBM RP3, with optical interconnects. Similarly, a paper by staff (Matsumoto, Sakano, Noguchi, Swabe) at NTT Transmission Systems Lab, 1-2356 Take, Yokosuka-Shi, Kahagawa-Ken, 238-03 Japan, described a system composed of 36 T800 Transputers running a parallel processing system with optical connections in "free space", i.e., in a box rather than in a chip. The technique looks promising but is also at a very early stage. A different view of optical devices was given by Dr. Peter Davis, who works at the ATR lab. See my report "The Advanced Telecommunication Research Institute (ATR) 14 June 1990." OPTICAL NEURAL COMPUTING One of the most interesting papers was on Mitsubishi's Optical Neural Neurocomputer. A paper was presented at this meeting and several other related papers have also been given at various neural network meetings during the summer months. The pricipal researchers are Kazuo Kyuma and J. Ohta at Mitsubishi Electric, Central Research Lab, 1-1 Tsukaguchi- honmachi 8-chome, Amagasaki, Hyogo 661 Japan. While there are many research activities, both in academia and industry, on the development of optical chips, Mitsubishi's looks very impressive to me because they have actually been able to fabricate a device that they claim can be mass produced. A important aspect of this chip is that there are only two layers rather than the three in other designs, by use of what they call a sensitivity-variable photodiode. Further the chip allows for a dynamic alteration in the interconnection weights between input and output. This last feature is essential for "learning", in neural networks, i.e., adjustment of interconnection weights in order to obtain specified output for given input. Ability to vary the weights is lacking in many other projects. The Mitsubishi group estimates that they can build this chip to contain about 1000 neurons in a one centimeter square with acceptable power requirements (less than one watt), large dynamic range (20db) for the weights, and low optical crosstalk. Although light beams do not singificantly interact at large scales, at chip level scales they do and this issue must be addressed. They estimate that such a chip can perform more than one tera connection per second, 10**(13). CONCLUSION. Optical computing appears to be a research area that has already provided a few practical applications. Its real potential for general computation lies several years away. There are many steps from discussion of AND gates to Fortran compilers. Nevertheless, this seems like a research topic with high payoff potential and only modest risk (cost). The Japanese government has set up an optoelectronics laboratory at Tsukuba and it would be natural for them to enlarge the scope of its research into optical computing, perhaps in conjunction with the activities that are present at the industrial labs. APPENDIX--EXPERT SUMMARY OF OC'90. Professor H.J. Caulfield Director, Center for Applied Optics University of Alabama in Huntsville Huntsville, Alabama 35899 Tel: (205) 895-603, Fax: (205) 895-6618 also attended this meeting and I asked him to provide me with a brief critical overview, which is given below. OVERVIEW OF THE KOBE MEETING, "OPTICAL COMPUTING '90" The Kobe, Japan meeting on Optical Computing was part of an annual series which is outlined below. 1985 USA 1986 Israel 1987 USAA 1988 France 1989 USA 1990 Japan 1991 USA 1992 Soviet Union 1993 USA 1994 Germany It is the one major international meeting in the field. The Kobe meeting was the largest so far - almost 600 attendees. This reviewer feels safe in saying there were no "breakthrough" papers. Accordingly, I offer this impressionistic assessment of work by regions. The European work is dominated by Germany, Israel, France, and the United Kingdom. In my judgement, the least dynamic of the groupings I have made. The Japanese work was short on innovation but long on practice. There were numerous papers on components they have manufactured especially Spatial Light Modulators (SLMs). The Seiko SLM was the most discussed. Most of the other Japanese work was on systems they had implemented. In general, they seem to be working on parts and system manufacturing methods, leaving the innovation primarily to the U.S. and the Soviet Union. I regard this as very clever on their part. Money and jobs stem from building things not inventing things. The Soviet Union was well represented - nearly all Russians. I was particularly impressed by their innovation. Papers V. Morozov's work on waveguide holograms coupled to diode lasers is symbolic of important work so far done only in Russia. American was well represented in number and breadth. Like the Soviet ork, the U.S. effort has no unified theme and even no agreements as to what is important. Alan Huang and many of his AT&T coworkers presented their well publicized but controversial (as to its current or potential value) work on digital optical computers. Optical neural networks, quantum mechanical effects, and hybrid analog-digital processors were other major themes. Overall, the field is growing in numbers and support, but no certain inner as a technology has emerged. The U.S. and the Soviet Union are the innovators. Europe is the monitor. These are caricatures, but they are not misleading. --------------------------END OF REPORT--------------
ishihara@etl.go.jp (Satoshi Ishihara) (09/22/90)
In article <7846@amelia.nas.nasa.gov> on 08/16/90 eugene@wilbur.nas.nasa.gov (Eugene N. Miya) wrote: |OPTICAL COMPUTING '90. I found and read Dr.Kahaner's report "optical computing in Japan" posted by eugene@wilbur.nas.nasa.gov to comp.arch and soc.culture. japan. I would like to thank them very much for their efforts to introduce our activity to the world. Just few remarks on the facts on our OC'90, now. (Sorry, but I have no time for detailed comments.) |A Proceedings, in English, is available from I don't know the definition of "Proceedings", but we call it "Conference Record", which collects 2-page summaries submitted for refereeing for the conference. |the International Society for Optical Computing xxxxxxxxx -----> Engineering Certainly, somebody wish to establish "the International Society for Optical Computing", but not yet:-) ^^^^^^^^^ They call the society SPIE. SPIE publishes it as Volume 1359 in their Proceeding Series. By the way, this "Conference Record" is also available at BCASJ (telephone: 03-817-5831) in Japan. ^^^^^^^^ Let me introduce a book with a similar title which was recently published in USA. ************************************************************** OPTICAL COMPUTING IN JAPAN Collection of about 50 papers. 525 pages, 1990. $85 (US/Canada) $98 (Other Countries-Includes Air Delivery). ISBN 0-941743-85-3. Contact: Nova Science Publishers, Inc. 283 Commack Road, Suite 300 Commack, New York 11725-3401, USA (516) 499-3103/3106 Telex #510 101 2161 Fax (1) 516-499-3146 *************************************************************** If there is anybody who is interested in the list of the titles of the papers with author's name, I will post it. ++ Satoshi ISHIHARA (Chairperson, OC'90 Steering Committee) Electrotechnical Laboratory(ETL), Japan e-mail: ishihara@etl.go.jp or ishihara%etl.go.jp@relay.cs.net