MLPerf
A set of industry benchmarks, run by the MLCommons group, that compares how fast systems train and run AI models.
Also: MLCommons
MLPerf began in 2018 and has been run since 2020 by MLCommons, an industry and academic consortium. It publishes results in rounds, for training, for data centre inference and for other settings, using fixed models and quality targets so results can be compared. Chipmakers, server makers and clouds submit; Nvidia and its partners submit in most rounds, and neoclouds such as CoreWeave, Nebius and Lambda have posted large-scale training results. Not every company takes part, and results reflect tuning effort as well as hardware.