Lessons from Loihi: Progress in Neuromorphic Computing
Mike Davies
Abstract
The past three years have seen significant progress in neuromorphic computing research, especially with Intel’s Loihi research chip enabling quantitative evaluation of algorithms and applications designed for this emerging computer architecture. These results have rigorously confirmed, for the first time, that significant gains in energy efficiency and latency are possible over a wide range of workloads compared to state-of-the-art conventional approaches. The greatest gains come from novel algorithms unrelated to the deep learning paradigm. While the speed, efficiency, and scalability of these algorithms suggest near-term commercial viability, Loihi’s high resource cost for large-scale workloads also highlights an urgent need for denser solutions for synaptic state in these architectures.