Rankings and lists, about 15 Jun 2027
MLPerf Training results
MLCommons' round of AI training benchmarks, where clouds show how fast thousands of GPUs train standard models.
MLPerf Training measures how long systems take to train fixed models to a set quality. Rounds are published in early summer and again in late autumn. Version 6.0, released on 16 June 2026, added a benchmark based on the 671-billion-parameter DeepSeek-V3 mixture-of-experts model, and the number of cloud systems submitted more than doubled from the previous round. Neoclouds use the results to show that their clusters scale: Nvidia's fastest result used 8,192 GPUs in GB300 NVL72 systems at CoreWeave. The date here is an estimate from past rounds.