MI300X vs H100 vs H200 Training Benchmarks

Scope and Summary A five-month, training-focused evaluation compared AMD’s MI300X against Nvidia’s H100 and H200 across GEMM microbenchmarks, single-node and multi-node training, collective communication, memory bandwidth, and performance-per-TCO. On-paper advantages for MI300X (TFLOPs, HBM bandwidth/capacity, Ethernet economics) did not translate into superior real-world training throughput on publicly available software. With hand-tuned developer images and work-in-progress (WIP) branches, MI300X improves materially,...