**2.2 Right time to revisit neuromorphic computing**

All current commercial accelerators for Artificial Intelligence (AI) computation employ the same computing paradigm that emerged more than 70 years ago based on the Turing computation model and the von Neumann stored-program computer architecture. However, these systems were not originally developed for targeting human-type sensory processing workloads that constitute majority of modern AI compute. This underscores the potential need to explore a much more complexityand energy-efficient computing model and architecture for AI computing. One approach is to revisit biology and examine how to mimic not just the functional behaviors, but also the structural organization of biological neural networks. Such an approach can potentially enable real intelligent computing with significantly less computation complexity and much better energy efficiency. One of the last major efforts taking such "neuromorphic computing" approach was by Carver Mead and his PhD students at CalTech back in the late 1980's utilizing analog VLSI circuits to model neural computation [4, 39, 40]. Both experimental neuroscience research and silicon fabrication technology have made tremendous advancements in the past 40 years, making this a good time to revisit the neuromorphic computing approach [6].

In recent years, several neuromorphic chips have been introduced both from academia and industry, including analog [41], digital [42–48] and mixed-signal [49] implementations. A dominant trend across these approaches is to implement a large number of spiking neurons communicating with each other via event-driven packets that encapsulate spiking information. Such spike-based event-driven computations have been shown to be more energy-efficient than traditional compute. However, they still lack the structural hierarchy and functional abstraction in the form of cortical columns as seen in the neocortex. Digital CMOS implementation of cortical columns and spiking neurons are both discussed as part of our research strategy in the next section.
