Evolving hardware on a large scale

Abstract

This paper presents a detailed technical description of a large-scale evolvable hardware system for evolving complex digital circuits directly in silicon at high speed. The core of the system is a three-dimensional array of reconfigurable logic with 5.2 million fine-grained function units and 1.2 Gbyte distributed memory. An application example is presented, describing an evolution of cellular automata based neural networks and a simulation of a hardware-based 75-million neuron artificial brain in real time. The system was developed in 1997-2000 at Genobyte, Inc. (Boulder, Colorado) for ATR HIP (Kyoto, Japan), and is marketed as CAM-Brain Machine (CBM). CBM features a true run-time logic reconfiguration, a hardware implementation of chromosome crossover and mutation, and a hardware-based fitness evaluation. CBM also features a sophisticated genotype-phenotype mapping through the process of embryonic growth.