Neural networks, which allow computers to perform pattern-recognition functions, have so far remained on the fringe of mainstream computing. But an agreement between Intel and Nestor to commercialize the Nestor-designed Ni1000 neural chip could change that. The Ni1000, which contains over 3 million transistors, can perform high-speed pattern recognition, from handprinted characters to fingerprints to images of targets in military applications. Nestor (Providence, RI) specializes in OCR and massively parallel-processing applications and is headed by Nobel prize laureate Dr. Leon Cooper.
According to Mark Holler, director of Intel's neural-network group, the Ni1000 chip performs 10 billion operations per second and is capable of recognizing 40,000 patterns per second, although other computer system bottlenecks reduce that number substantially. Holler says that some prototype systems using the chip and Nestor's OCR software, NestorReader, can recognize up to a few hundred handprinted characters per second.
Dr. Don Specht, consulting scientist at the Lockheed Research Laboratory, a beta site for the Ni1000, says that Lockheed is working on target-recognition applications using four Ni1000 chips running on a VME board and has plans for additional space-based and commercial applications. Intel's Holler says that there are eight beta sites working with the Ni1000. He sees a wide variety of industries making use of neural-network technology (see the figure).
Intel is working on reducing the cost of the silicon for the Ni1000, says Holler (pricing at press time was unavailable), and eventually, it will release an add-in card for PCs similar in cost to a "typical high-end graphics accelerator." Intel is supporting the chip with a C library and other software development tools, in addition to the NestorReader OCR software.
While neural networks are still a specialized segment of computing, commercial efforts such as Intel's and Nestor's are bound to make their application more widespread. The chip is expected to ship sometime in the first quarter of this year.
For more information, call Intel's Neural Network Hotline at (408) 765-9235.