Hardware Support for Distributed Associative Memories

This paper describes a hardware to support distributed associative memories, in particular the ADAM neural network architecture. The system is designed to support a large number of moderately sized ADAM memories on a distributed memory MIMD architecture. The system addresses the issues of weight paging to allow neural networks of very large size to be executed using only a relatively small physical memory. The hardware also supports dedicated peripheral logic to accelerate the most compute intensive aspects of the ADAM memory. The paper describes the benefits gained by implementing the ADAM in hardware and presents a discussion of its design and use.