HAVENN: horizontally and vertically expandable neural networks
Jien-Chung Lo, Georg Fischer · 2002
The toughest challenge facing hardware designers of artificial neural networks is the expandability problem, since no single VLSI chip is likely to accommodate all components of a real world application. In this paper, the authors present a microelectronic system architecture with virtually unlimited expandability at a relatively low cost in additional hardware and reduced system performance. The horizontally and vertically expandable neural network (HAVENN) architecture consists of three types of chips: a single layer neural network chip, a summer chip and a repeater chip. The most important features of the proposed architecture are: a balanced distribution of(circuit) complexity between board level and chip level, easy implementation, true parallel operation and versatility.>