Neural network chips from Intel and Adaptive Solutions

In 1989-1990 two chips appeared that were made for neural networks. Intel's analog ETANN had 64 neurons and 10,240 synapses, its weights held as charge on floating gates, and computed more than 1.3 billion connections a second. Adaptive Solutions' digital CNAPS had 64 processors at 25 MHz and learned on the chip itself.

ETANN, by its Intel designers' 1993 abstract: a 1-micron CMOS non-volatile memory process, a 64 x 64 EEPROM weight array, more than 1.3 billion connections a second, and more than 2 billion in a 2:1 multiplex mode. Users of the chip in 1992 (an AFIT thesis working from Intel's manual, and a Fermilab report) write that weights are retained for at least 10 years and that learning is mostly done off the chip; one gives a weight resolution of 7.5 bits, the other about 6. CNAPS, by Adaptive Solutions' NIPS 1990 paper: all processors receive the same instruction, 1.6 billion inner product steps a second per chip, a 9 x 16-bit multiplier, 16-bit weights. An eight-chip board updates 2.3 billion connections a second in learning and processes 9.6 billion in the forward pass, but these figures come from a clock-accurate simulator, not measured silicon. What the record does not claim: that these were the first commercial neural network chips, or the month ETANN was shown; neither the 1989 paper nor Intel's data sheet could be read.