The Silicon Optic Nerve

Abstract

Communication among neuronal elements is a significant limiting factor in the design of VLSI neuromorphic systems. This fact is not surprising given that a large percentage of the volume of the nervous system is composed of myelinated axons. The degree of convergence and divergence of single neurons is staggering in comparison with human-made computers. It might appear impossible, even in principle, to build such structures in VLSI circuits, which are limited to a two-dimensional plane of silicon. Surprisingly, the cortices of the brain are nearly two-dimensional as well. In fact, we cannot increase the degree of connectivity in a system whose wires occupy space by employing a structure in which nodes are arrayed in three dimensions [97]. There is nothing fundamental about the structure of neural tissue that cannot be embedded in silicon. The thickness of cortical structures can be represented with a correspondingly larger silicon surface area. However, silicon surface area is available on small die, which are several millimeters on a side, and so the number of neurons that can be fabricated on a single die is limited. Consequently, connections between silicon neurons located on different chips are essential for building even moderately sized artificial neural systems. This chapter introduces a general method for interchip communication that is tailored for use in neuromorphic systems.