Neuromorphic research aims to approach the capabilities and efficiency of the brain by developing algorithm and hardware architectures which adopt mechanical features of biological computing. As improvements in computing speeds slow and new workloads scale towards untenable costs in conventional approaches, neuromorphic computing offers a novel paradigm for neural network intelligence and hardware architectures which may realize a next-generation of computing performance. Such a paradigm optimizes for different goals than conventional computing systems, and therefore novel benchmark assessment and methods are necessary for technological progress. In this article, we present NeuroBench, a benchmark framework for neuromorphic computing which has engaged over 100 researchers from more than 50 institutions across academia and industry. NeuroBench provides novel benchmark tasks and common, extensible tools for evaluation, unifying the space of neuromorphic research. Results from NeuroBench benchmarks and challenges highlight the strengths of neuromorphic algorithms and systems for sparse, event-driven, and energy-efficient machine learning.