Super chips for artificial intelligence

There is at least one field - artificial intelligence - that can effectively utilize air of the projected advances of semiconductor technology for the foreseeable future. The potential applications will be just as widespread, with education, health, industry, and service sectors among the many areas leading to a large demand for highly sophisticated solid state components. This address will assess these expanding virtues and their effect on the IC. Currently prevailing views about the future demand for high-powered computational elements are sometimes misguided. One such myth says that the demand for 32b (or larger) processors is likely to be limited because most applications do not need so much computational power. What these estimates fail to take into account are the computational needs of error-tolerance and user-friendliness that are essential if we expect ordinary mortals to use tomorrow´s microcomputers effectively. The second myth stems from the view that a high powered LISP computer will be adequate to handle all of AI´s computing needs. There are many topics that arise in AI for which the LISP machines alone are not adequate: perception problems for vision and speech, motor problems of manipulation and mobility in robotics, symbol manipulation problems of language, repetitive knowledge, recognition of rule-based inference in expert systems, combinatorial search in game playing and puzzle solving, massively parallel processing for instantaneous extraction and matching of relevant information from large knowledge bases. Each of these problems involves a wide variety of computational tasks, making it difficult to select any one architecture as the most desirable. A more probable AI architecture would be a multiprocessor structure with a large number of asymmetric processing elements sharing a large knowledge base. Furthermore, the trend from VLSI to ULSI will be towards higher degrees of Integration leading to higher functionality. A million-element chip m- ght include functions currently requiring four or five chips. CPU, virtual memory, floating point arithmetic, array processor and cache. A hundred-million element chip might contain a 100 processor system with a common bus to a larger external shared memory. It is important to note that all of the hundred processors need not be identical. Processing elements should have dynamically alterable PLAs, leading to tailored architectures for different computational tasks. A billion-element super-chip might provide a 100 megabyte common memory and 100 processing elements on a single chip, along with A/D and D/A and networking interfaces. Even a super chip could barely perform many of these AI problems. A review of the foregoing problems illustrates the need for flexible and powerful computational elements in the future. It is indeed possible that we can haven billion element super chip by the turn of the century: it is interesting to speculate how we might use it!