Hardware architectures and implementations for associative memories---the building blocks of hierarchically distributed memories

During the past several decades, the semiconductor industry has grown into a global industry with revenues around $300 billion. Intel no longer relies on only transistor scaling for higher CPU performance, but instead, focuses more on multiple cores on a single die. It has been projected that in 2016 most CMOS circuits will be manufactured with 22 nm process. The CMOS circuits will have a large number of defects. Especially when the transistor goes below sub-micron, the original deterministic circuits will start having probabilistic characteristics. Hence, it would be challenging to map traditional computational models onto probabilistic circuits, suggesting a need for fault-tolerant computational algorithms. Biologically inspired algorithms, or associative memories (AMs)—the building blocks of cortical hierarchically distributed memories (HDMs) discussed in this dissertation, exhibit a remarkable match to the nano-scale electronics, besides having great fault-tolerance ability. Research on the potential mapping of the HDM onto CMOL (hybrid CMOS/nanoelectronic circuits) nanogrids provides useful insight into the development of non-von Neumann neuromorphic architectures and semiconductor industry. In this dissertation, we investigated the implementations of AMs on different hardware platforms, including microprocessor based personal computer (PC), PC cluster, field programmable gate arrays (FPGA), CMOS, and CMOL nanogrids. We studied two types of neural associative memory models, with and without temporal information. In this research, we first decomposed the computational models into basic and common operations, such as matrix-vector inner-product and k-winners-take-all (k-WTA). We then analyzed the baseline performance/price ratio of implementing the AMs with a PC. We continued with a similar performance/price analysis of the implementations on more parallel hardware platforms, such as PC cluster and FPGA. However, the majority of the research emphasized on the implementations with all digital and mixed-signal full-custom CMOS and CMOL nanogrids. In this dissertation, we draw the conclusion that the mixed-signal CMOL nanogrids exhibit the best performance/price ratio over other hardware platforms. We also highlighted some of the trade-offs between dedicated and virtualized hardware circuits for the HDM models. A simple time-multiplexing scheme for the digital CMOS implementations can achieve comparable throughput as the mixed-signal CMOL nanogrids.