Bio-Accelerators: Bridging Biology and Silicon for General-Purpose Computing

As the historical trend of speed and energy efficiency improvements diminishes [2], radical departures from conventional approaches are becoming critical to improving the performance and energy efficiency of general-purpose processors. Inspired by biological nervous systems, neuromorphic computing delivers high performance within a very low power envelope. In fact, human brains can use only 20 Watts to carry out tasks which require a warehouse of processors to accomplish. Despite these advantages, neuromorphic computers are not easily programmable in the way that traditional von Neumann machines are. Furthermore, such models have inherent inaccuracy in their computation, which must be addressed by the programer. Towards reconciling these differences, our past research has focused on bridging neuromorphic and von Neumann computing models through intuitive programming models and architectural interfaces, while embracing the notion of approximate execution [3]. Of course, this work has only used artificial digital or analog neural networks. We seem to have forgotten the most powerful neural network of all — a biological neural network! We imagine a future technology in which computational neuroscience and computer architecture intersect, leading to a programming environment which offloads approximable regions of code onto the very brain of their users. In this framework, the biological nervous tissue becomes an accelerator for a code written in conventional programming languages. We refer to these accelerators as bio-accelerators, and explore their function, strengths and limitations.