Mead names neuromorphic systems
In October 1990 Carver Mead of the California Institute of Technology called analog chips built on the organising principles of the nervous system neuromorphic systems. His example was the silicon retina he had described with Misha Mahowald: about 100,000 devices doing 100 million operations a second on 1 mW, about 10^-11 J an operation against 10^-7 J in a digital design.
Why it matters
The paper named a line that looks for efficiency not in a faster digital processor but in analog physics close to nervous tissue. The comparison of energy per operation it opens with is still that line's main argument.
By the paper, a microprocessor spends about 10^-7 J on one operation, and the brain is a billion times more efficient than the digital technology of the day. The retina, Mead writes, is still five orders of magnitude behind the brain but four ahead of digital methods; the paper concludes that adaptive analog systems use 10,000 times less power than comparable digital ones. The group had fabricated hundreds of test chips through DARPA's MOSIS service. The abstract of the 1988 Neural Networks paper says an analog model of the first stages of retinal processing was built on a single chip: each photoreceptor computes the logarithm of the light intensity, a resistive network smooths the signals, and measured outputs from an experimental 48 x 48 pixel array show many properties of the bipolar cells of the vertebrate retina. What the record does not claim: that the 1990 paper was the first to use the word neuromorphic, which it does not say; or the brain's energy per operation, which the scan prints without a minus sign and so was not taken. The full text of the 1988 paper was not read.