Meta Reports Up to 1.5x Faster AI Training on AMD GPUs

Summary

Meta describes an end-to-end approach for running large-scale AI training across a heterogeneous fleet of NVIDIA GPUs, AMD GPUs, and proprietary MTIA accelerators. The paper focuses on recommendation-model training with AMD Instinct MI350X and NVIDIA B200, arguing that performance portability requires co-designing the model and software for each accelerator instead of directly porting NVIDIA-optimized implementations. The authors highlight MI350X's 1.6x higher HBM capacity relative to B200 as one example of a hardware characteristic that can shape model and training choices. They also introduce an analytical performance model that selects kernels without a warm-up phase across the datatypes and matrix shapes used by production models, reducing manual tuning. In addition, Meta extends agentic workflows for CUDA-to-AMD migration with microarchitecture-aware kernel optimization and custom shim generation. Together, these methods improve end-to-end training throughput by up to 1.5x over a directly ported baseline. The paper presents the work as the first hyperscaler report on productionizing AMD GPUs for large-scale training and says the resulting contributions have been upstreamed to PyTorch, Triton, and ROCm.