Meta trains recommendation models on AMD GPUs in production

Meta published a paper (arXiv:2610.22230, accepted at MICRO 2026) on industrial training of recommendation models on AMD Instinct MI350X versus NVIDIA B200, where co-design of models and hardware, analytical kernel selection without warmup, and agentic code migration from CUDA yield up to 1.5x higher training throughput compared to direct porting.

Compiled by Sergey KostenchukPublished 2026-10-19Updated 2026-10-19

2026-10-19 Coding Meta

💻 Meta trains recommendation models on AMD GPUs in production

Meta published a paper (arXiv:2610.22230, accepted at the MICRO 2026 conference). It trains recommendation models on AMD Instinct MI350X versus NVIDIA B200: co-design of models and hardware (MI350X has 1.6x more HBM), analytical kernel selection without warmup, and agentic code migration from CUDA yield up to 1.5x higher training throughput compared to direct porting. Changes have been contributed to open-source PyTorch, Triton, and ROCm.

🌍 The first hyperscaler report on training on AMD GPUs reduces the industry's dependence on CUDA and provides patterns for heterogeneous NVIDIA/AMD/MTIA fleets.

👤 Useful for those building ML infrastructure: the methodology is transferable to other hardware, but the exact 1.5x gain depends on Meta's proprietary pipelines.

Source 1: https://arxiv.org/abs/2610.22230 Source 2: https://news.ycombinator.com/item?id=50147458

Sources