rick@cs.arizona.edu (Rick Schlichting) (05/21/91)
[Dr. David Kahaner is a numerical analyst visiting Japan for two-years under the auspices of the Office of Naval Research-Asia (ONR/Asia). 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. All information is dated and of limited life time. This disclaimer should be noted on ANY attribution.] [Copies of previous reports written by Kahaner can be obtained from host cs.arizona.edu using anonymous FTP.] To: Distribution From: David K. Kahaner ONR Asia [kahaner@xroads.cc.u-tokyo.ac.jp] Re: ICOT 17 May 1991 ABSTRACT. A summary of Japan's Fifth Generation Computer Project is given. INTRODUCTION. Beginning in 1982 Japan embarked on a ten year program called Fifth Generation Computer Project (FGCP), funded at over 40billion Yen from government sources. The goals of the project were to focus on large-scale parallel-processing for non-numerical computation, in particular on logic programming. The project generated a great deal of excitement in the West, partially because of the implications for Japan of its success, and partially because of questions about its scientific feasibility. The program is managed and most of the research is performed at the Institute for New Generation Computer Technology (ICOT) Mita Kokusai Bldg-21 Floor 1-4-28 Mita, Minato-ku Tokyo 108 Japan Tel: +81-3-3456-3195, Fax: +81-3-3456-1618 (Many ICOT staff have electronic mail access. One such is, Dr. Koichi Furukawa, Research Center Deputy Director, FURUKAWA@ICOT.OR.JP.) I have deliberately not written much about ICOT. Because it has been in existance since the early 1980s I felt it was well known already, especially since much of its output is in English. Nevertheless as the program gets near its completion it is appropriate to provide a little summary material. The ICOT Journal (English) is extremely well written and contains research papers and many references. In particular, Volume 29, 1991 contains summaries of the 1990 FGCP Symposium, held June 1990. Its contents are worth describing in some detail. Opening Address. Hirochi Hiroshige (Executive Director ICOT) Brief summary of reports to follow at the Symposium. Greetings. Yoshihiro Sakamoto (Dep Dir Mach and Info Indus Bureau, MITI) How Can the Fifth Generation Computer Technologies Contribute to the World Today? Hajime Karatsu (Tokai University) Primarily an exposition of why Japan has been so successful, comments about the role of manufacturing, and industry projections. About FGCP, mostly remarks about the friction it caused in the West, why Western thinking about the project was incorrect, and the need to apply FGCP results to improve productivity, especially for older people (Japan will have many of these in future). Future Development of Parallel Inference Technology. Kazuhiro Fuchi (Dir ICOT Research Center) A history of the software and tools activities, mostly associated with logic programming, guarded horn clauses, use of KL1 as language. Overview of the Fifth Generation Computer Research and Development Project and Research Organization. Takashi Kurozumi (Dep Dir ICOT Research Center) Detailed charts of organizations involved in FGCP, overseas activities, hardware and software developments, policies and themes. Summary of Research and Development in Fiscal 1989 and Results Achieved. Shunichi Uchida (Manager, Research Dept ICOT Research Center) More details about various research activities, experimental programs during the past year. What Will Parallel Programming be Like in the Future? Panel Discussion. Koichi Furukawa (Dep Dir Research Center ICOT) Toshinori Watanabe (Hitachi) Hiroshi Okuno (NTT) Ryuzo Hasegawa (ICOT) Yuji Matsumoto (Kyoto University) Kazuo Taki (ICOT) Very interesting discussion, especially when contrasted with similar panels in the US which are composed primarily of people concerned with numerical applications. Various models of PIM (Parallel Inference Machine) have been built as protoyptes. These are summarized in the chart given below. (PIM/p is the latest.) PIM/p PIM/c PIM/m PIM/i PIM/K Machine RISC+macro Horiz Horiz RISC RISC Instructions micro micro Target Cycle 50 ns 50 ns 50-60ns 100ns 100ns Time LSI Devices Standard Gate Cell base Stand Custom cell array cell Process Tech About 1mu-m 0.8mu-m 0.8mu-m 1.2mu-m 1.2mu-m (Line width) Machine Multicluster Multiclust 2-Dim Shar mem Two Configuartion connections in X-bar mesh thru level in hyper^3 (8 PEs+CC parallel par (8 PEs linked linked to cache cache to shared mem) shar mem) Number of PEs 512 256 256 8PEx2 16PEx2 Various application programs have been written on PIM, including VLSI (logic design, routing, logic simulation), Go playing, Genetic information processing, Legal reasoning, Parallel constraint solving, Parallel theorem proving, and Parallel parsing. In Tokyo, the US Embassy's Office of S&T Affairs (Science Office) has prepared a summary of FGCP. The summary was prepared by scientists (not computer scientists) and written for other non-specialists. I was able to read and comment on a draft of this summary, and I believe that it is as accurate a description of the project as is possible given its intended audience. At the end of this summary there are some remarks about the New Information Processing Technology project, set to begin in April 1992. For more details about that I refer readers to my reports (nipt.90, 26 Dec 1990) and (nipt391, 7 April 1991). 1. SUMMARY The MITI directed ten-year program of the Institute for New Generation Computer Technology (ICOT) to develop a Fifth Generation Computer System (FGCS) ends on March 31, 1992. During 1-5 June, 1992, ICOT will host its fourth and final conference where the project's final results will be presented. The project has met most of its goals, albeit on a small scale. Parallel Programming Languages have been developed which are no more difficult to use than are current sequential languages. ICOT computers do make inferences, but from small knowledge bases. The computers do not use spoken languages in their operation, nor do they translate between languages. Neither is the speed of the ICOT computers as fast as that of numerical computers. 2. The future looks hopeful, though, for the ICOT computers. ICOT and U.S. researchers have identified several applications for the computers (e.g., expert decision systems, legal reasoning programs, genetic sequencing) which may be realized during the 1990s. Also, MITI may continue supporting one or more parts of the project past next year; the Ministry's decision may come as early as this fall. 3. MITI is now deciding whether to launch a subsequent computer initiative, the New Information Processing (NIPT) project. If feasibility study results turn out positive, the new program will begin in April 1992. End Summary BACKGROUND 4. In 1979, MITI assigned the Electrotechnical Lab in Tsukuba the task of defining a project to develop a computer system for the 1990s. MITI accepted the early reports on the project, presented its plan for the project at an international conference in October 1981, and, in April 1982, initiated a ten-year Fifth Generation Computer Systems (FGCS) project establishing the institute for Next Generation Computer Technology (ICOT). 5. ICOT is a consortium of Japanese companies - including associate members IBM Japan, DEC Japan, and UNISYS Japan - and MITI Institutes with a Central Research Lab. The number of staff at the Lab has ranged from 50 initially to about 90 now; they stay in ICOT for 2-4 years and then rotate back to their companies' research institutes. Fewer than a dozen people have been with ICOT from its beginning. ICOT R&D is funded by MITI. Annual funding for the project has risen from less than 3 billion yen to nearly seven billion yen. This figure includes salaries of the 90 staff members but does not include management and administration services - provided by NTT - and building rental fees; these are paid for by donations from industry which annually total about 500 million yen. 6. ICOT's main activities consist of developing a hardware and software prototype system, conducting joint research at overseas research centers, and studying new information processing technologies and new generation computer applications. Dr. Kazuhiro Fuchi, ICOT Managing Director, told the Science Office staff that ICOT's goal has been to develop an easier-to-use non-Von Neumann Computer System with parallel processing ability and an ability to make inferences when executing a program. The key words, he said, were "parallel" and "inference." To accomplish this, ICOT has been developing hardware and software for each of three subsystems - knowledge base, problem solving and inference, and human-machine interaction. Programming is done "in house" as is computer design; hardware manufacture is contracted out. Contrast with Earlier Computers 7. According to Takashi Kurozumi, Deputy Director of ICOT's Research Center, the title "Fifth Generation" was used to differentiate the project from the earlier generations of computers which had been classified according to their constituent hardware elements developed to increase storage: vacuum tubes; transistors; integrated circuits; and LSI and VLSI chips. These kinds of computers are characterized by addressable memory or stored program schemes, sequential processing, and numerical calculation. 8. Shunichi Uchida, the Manager of ICOT's Research Department, described the development of such Von Neumann computer systems as follows: First the hardware is designed, then the software, then the applications are devised. He said the development of the ICOT computer would follow a different path: First concept, then language/software, then hardware, then applications based on parallel knowledge processing. ICOT's computer was to be based on a new framework, a framework that would allow the computer to do processing that depends on circumstances, more in the style of human thought, rather than having to follow pre- defined procedures. Uchida said the power of the ICOT computer is measured in logical inferences per second (LIPS) rather than in instructions per second (IPS). At the Beginning 9. According to documents published in the early 1980's, the project goal was very ambitious. It was to develop a knowledge information processing system (KIPS) - something like an expert system - with an intelligent conversation function and an inferential function employing a knowledge base. The prototype computer would acquire, accumulate, and use various types of knowledge; it would infer information from what previously was not known explicitly; it would make conjectures based on incomplete knowledge; it would use natural written and spoken language, graphics, and other types of image input data; it would effectively translate between languages at a semantic level. The prototype computer would use a new programming language based on new principles thereby enabling people with no expert knowledge to write programs with ease. At the 1988 FGCS Conference, ICOT Research Center Deputy Director Takashi Kurozumi reconfirmed most of these goals. He did not include the language translation goal nor the goal of using spoken language in communication with the computer. 10. To succeed, the project needed to make major technological innovations to enable the fifth generation computer to support very large knowledge bases, allow very fast associative retrievals, and to perform logical inference operations as fast as other computers perform arithmetic operations. The project also needed to utilize parallelism in program structure and hardware to achieve high speed, and to develop a machine-user interface allowing significant use of natural speech and images. ICOT was not to be an artificial intelligence (AI) project, but ICOT researchers hoped to combine their own research with that being done in artificial intelligence, and in architecture and software technology. Fuchi and Uchida, who have been with ICOT since 1982, told the Science Office staff in March 1991 that the project has not included VLSI and device techniques nor has it included neural networks. 11. One of the FGCS's significant characteristics was the decision to use the European programming language Prolog as the machine language of the logic processor. Prolog is an "if...then" language, or a predicate logic language, not designed for numerical calculations. It is useful for situations and ambiguous activities where the procedures that the computer is to perform cannot be clearly stated in the program at the outset, where the way to proceed is not explicitly specified. That was done, Fuchi said, because the project was to develop technology to carry out high-level symbolic operations, not numerical calculations. One American computer expert, evaluating prolog, said it represents knowledge elegantly but often opaquely and arcanely. He stated further that Prolog solves automatically the problems it sees, without the user being involved. This is not universally considered a plus, he noted, because knowledge engineers typically do not want to abdicate step-by-step control to a process that conducts massive searches automatically. Results 12. Because this was Japan's first attempt at parallel computing, ICOT began by developing sequential inference machines. Development continued throughout the project, culminating in the completion this year of the PSI III machine having 1.2 million lips of power. (ICOT officials consider 10-l5 million LIPS to be equivalent to 400 MIPs.) From this base, multiple processor sequential inference machines were built, as were components for a parallel inference machine. The 1992 hardware goal is a 1,000 processor parallel inference machine (PIM) having 300-600K LIPS Power. A 32-processor parallel inference machine has been built, and the larger machine will have 31 additional 32-processor elements. 13. In 1989, the Japanese brought ICOT's sequential inference machine PSI II to the United States. It generated interest and resulted in a visit to Japan in February 1990 from Argonne National Laboratory of scientists to discuss cooperative activities. This past September, ICOT also hosted a bilateral workshop on parallel knowledge systems and logic programming; more than a dozen participants came from the U.S. side alone. Argonne has been doing logic programming and genetic information processing research, and the lab now has two of ICOT's PSI II machines (power: 330K LIPS) which are being evaluated. Argonne will host an NSF- Sponsored workshop this June on parallel theorem proving. NIH officials are actively considering using ICOT's machine for biological work. The Lawrence Berkeley Lab has network access to ICOT, and ICOT is building its own 64KBIT line in Japan to increase domestic access to its computers. (Currently, for example, Tokyo University cannot access ICOT's computers.) 14. ICOT officials have wanted to provide programmers with a parallel system for which they can write programs as easily as they can write programs for Von Neumann Computers. Project researchers, therefore, have developed several parallel programming languages for the computers they have developed. ICOT's longest logic language program, KL-1, is 100,000 lines and growing. This, Uchida said, was the world's longest program written for parallel processors and was equivalent to more than 300,000 lines of a Von Neumann language program. In addition, ICOT has developed the world's only full-scale logic language operating systems - SIMPOS, for its sequential machines, and PIMOS for its parallel ones. Finally, the latest versions of the ICOT machines also have some Unix capability. 15. ICOT programs can work on other parallel computers, said Uchida, albeit much slower than on ICOT computers, in part because other parallel computers were designed to accommodate Von Neumann programming for numerical computation. The front end processor of such a computer is designed to find parts of such linear programs that can be worked in parallel, to send those parts out to the various processors, and then to reintegrate those parts back into the program. ICOT'S non-numerical programs do not run efficiently on such machines. On the other hand, Fuchi noted, since Cray (and other parallel) computers were designed (for numerical computation and thus) to accept languages such as Fortran there is much more demand for them today than there is for ICOT's non-numerical computers - which cannot use Fortran or other sequential programming languages. 16. There have been three conferences held in connection with the project -in 1981, 1984, 1988. ICOT will sponsor the project's final conference in the Tokyo Prince Hotel on 1-5 June 1992, where final results will be presented. ICOT publishes a quarterly journal which is distributed to over 1,100 locations - 600 of these in 36 foreign countries. ICOT also receives long-term researchers from the United States, the U.K., and France, and short-term researchers from those and a variety of other countries. Dr. Mark E. Stickel, SRI International, currently is being supported at ICOT through an NSF award. Dozens of personnel sequence inference machines have been commercialized by the companies party to ICOT. 17. The project has met most of its goals, albeit on a very small scale. Parallel programming languages have been developed which are no more difficult to use than current sequential languages. ICOT computers do make inferences, but from small knowledge bases. The computers do not use spoken languages in their operation, nor do they translate between languages. Fuchi told the Science Office staff that efforts have not been made toward voice recognition because such a computer would be a "special" computer, and ICOT's computer was to be a "general, all- purpose" computer. The computer's speed also is not as fast as that of numerical computers. 18. Uchida indicated several future applications for a parallel inference system. These included VLSI CAD applications, theoretical math systems, a variety of expert decision systems, legal reasoning and grammar/syntax programs, genetic sequencing and other applications in biology, and go-playing programs - similar to current chess-playing programs. He did not speculate, though, on when these applications might be realized. 19. Fuchi told the Science Office staff that ICOT staff members are currently reviewing the project to decide which parts of it should continue past next year; he expected an announcement on this to be made by the end of the summer. In his opinion, the most successful part of the project is the hardware model based on the PIMOS operating system. Fuchi thought that the world's computer industry is "not yet mature enough to capitalize on the research coming out of ICOT," that ICOT was "too far out in front" of the rest of the computer industry. He said he will suggest to MITI that it will be the Government's responsibility to nurture the ICOT computer until industry is ready for it. 20. ICOT's future also might be spelled NIPT. The Embassy reported on the new information processing technology (NIPT) initiative last year (Reftel A) after the December International Workshop on the project and last month (Reftel B) after attending that NIPT International Symposium on March 13-14. MITI officials plan for NIPT to succeed ICOT in 1992. The NIPT project differs in several respects with ICOT, however. First, NIPS is from the beginning an international project; MITI officials have been actively seeking international input into the design of the NIPT program. ICOT has been a domestically-planned MITI project, but an open one whose information has been shared with the rest of the world. 21. SECOND, NIPT is building on the advances made during the 1980s by ICOT and others in computers and associated technology, although at the moment MITI officials do not see NIPT as directly incorporating any ICOT technology. NIPT probably will concentrate on massively parallel systems which will be optically connected. It will benefit from advances made in VLSI and VLSI technology. Advocates hope they can use future advances in biology, physiology, psychology, and neurology to develop neural networks for the NIPT computer. Third, NIPS' initial focus is principally on hardware; ICOT's principal initial focus was more on software. Fourth, NIPT computers will be able to crunch numbers; ICOT's cannot. Fifth, the NIPT project will not be conducted in a single institute, such as ICOT, and it will have many R&D objectives. 22. NIPT will be funded initially at a lower level than ICOT currently is. MITI officials have indicated that NIPT will be funded at 30 million dollars per year, after the feasibility study. This is about 60 percent of current ICOT funding. There is much more interest in NIPT worldwide, especially within industry and government, than there has been in ICOT which has generated almost all of its interest in academia. Nonetheless, ICOT has been a respectable program; its researchers have made major developments in inference languages and consideration is being given to tie-ups with vector supercomputers, such as at the Pittsburgh Supercomputer center in the United States. Its latest machines are becoming more Unix-capable. If the GOJ continues to support the program after 1992, if an increasing number of applications can be found for ICOT's computers, and if users can be convinced that there is money to be made by these non-numeric applications, ICOT R&D may continue in parallel with NIPT. How appropriate. -----------------END OF REPORT-----------------------------------------