MlPerf Training v2.1 Results New Nvidia press releases and ML Perf benchmarks also seem more and more hypocritical each time. Notable: - No consumer GPUs at all - Highly sublinear scaling for many tasks (4-8 GPUs seem most efficient) - No proper apples-to-apples H100 test - AMD EPYC Vendor lock in place - "Open" tests are barely there - As predicted this is a playground for hardware vendors - On RNNT and Imagenet H100 2x faster, on BERT - 3-4x