An 11-million Transistor Neural Network Execution Engine

This 11-million transistor, 25MHz neural network execution engine combines digital circuit technology with proven redundancy techniques to produce a chip capable of performing 1.6 billion connections/s. The device measures 26.2x27.5mm and is fabricated in a 0.8µm CMOS process. (Table 1) The processor array is composed of 64 processor nodes (PNs). 16 PN spares are included to maximize yield.1 Single clock execution, large embedded memory,and an optiinized single-instruction multiple-data (SIMD) architecture are key to performance.2 This single-chip implementation of the computing power required for neural alogrithms appeals to a broad range of applications.