RA: an analog neurocomputer for the synchronous Boltzmann machine

The Random Analog (RA) Neurocomputer runs synchronous Boltzmann machines dedicated to real time visual pattern recognition. Its architecture achieves scalability and modularity thanks to a linear array of analog processing modules linked by a high bandwidth interconnection network, a controller, and a front end running a high level user interface. A prototype with 128 input neurons and 64 active neurons and 4.2 GCPS peak processing power is described and simulation results are given.