Multifunction content addressable memory for parallel speech understanding

Abstract

Content Addressable Memories (CAMs) allow considerably finer-grained parallelism than conventional shared or distributed memory multi-processors. This fine-grained "Processor-In-Memory" concept can be employed to a large degree during Semantic Network processing in support of Artificial Intelligence (AI) with specific applications in speech and natural language processing. A special-purpose CAM configuration is presented based on requirements for a nominally-sized 64 K node semantic network with 8 bit-markers and 32 relationship types. Analysis for a target application shows that the extensive use of parallel Marker-Propagation and Set Theoretic Operations yields approximately 30-fold speedup over systems with standard Random Access Memories.