Focus on benchmarks for neuromorphic computing

Neuromorphic (brain-inspired) computing technology has been of interest to researchers since its origins in the work by a group at CalTech led by Carver Mead in the 1980s. More recently this interest has extended into the commercial domain with major industrial players such as IBM and Intel exploring the technology, and start-up companies commercializing neuromorphic solutions to applications such as inference in low-power edge systems through to datacentre-scale alternatives to GPUs for large language models and deep learning. With this growing commercial interest it is increasingly important to be able to compare and contrast the strengths and weaknesses of alternative neuromorphic offerings that range from the sub-threshold analogue circuits favoured by Mead's seminal work through novel device technologies such as memristors that offer physical in-memory compute capabilities, all the way up to large-scale many-core digital systems based upon conventional (and highly manufacturable) digital technologies. Such comparisons require benchmarks as the basis for comparison, but the sheer diversity of current neuromorphic technologies creates difficulties for prospective benchmarks. This Focus Issue aims to pull together some early thinking on neuromorphic benchmarking. This comes in various forms, including comparing the same application on two different neuromorphic platforms and seeing which applications demonstrate a neuromorphic advantage over conventional solutions. The collected papers represent early perspectives on the neuromorphic benchmarking challenge but they are far from the last words on the matter—there is still a great deal left to do here!