Meta details how it made AMD GPUs production-ready for training
Porting recommendation models from Nvidia to AMD took more than a driver swap
The facts
- Meta used AMD Instinct MI350X versus Nvidia B200 for recommendation model training as its case study
- Meta reports up to 1.5x improvement in end-to-end training throughput versus a directly ported baseline
- The paper cites MI350X's 1.6x HBM (high-bandwidth memory) capacity relative to B200 as one exploitable hardware difference
- Meta says this is the first hyperscaler report on productionizing AMD GPUs for large-scale training
- Contributions have been upstreamed to PyTorch, Triton, and ROCm
- Work is co-authored by 58 people and will appear at MICRO 2026
Meta has published a detailed account of getting AMD's Instinct MI350X GPUs to run recommendation-model training at production quality. The goal was not simply porting code written for Nvidia hardware. The paper, co-authored by 58 Meta engineers and due at the IEEE/ACM MICRO 2026 conference, says a straight port from Nvidia leaves performance on the table. Years of CUDA-specific optimisation do not transfer cleanly to a different architecture.
The company's fix is three-part. It redesigned model structures and software recipes to exploit AMD-specific traits, rather than treating the chip as a drop-in replacement. One example is the MI350X's 1.6x HBM (high-bandwidth memory) capacity compared with Nvidia's B200. It also built an analytical performance model that picks the optimal compute kernel for a given datatype and matrix shape. This avoids the warm-up tuning runs such selection normally requires. And it extended agentic coding workflows, AI systems that write and test code with little human intervention, to handle microarchitecture-aware kernel optimisation. Those workflows also generate the shim code needed to bridge the CUDA ecosystem to AMD's ROCm stack.
Meta reports that these methods lift end-to-end training throughput by up to 1.5x. The baseline for that comparison was a version ported directly, without the co-design work. The company says it has upstreamed its contributions into PyTorch, Triton and ROCm. Other developers using those open-source projects should see at least some of the benefit without reimplementing Meta's work themselves. Meta describes the paper as the first report from a hyperscaler on productionising AMD GPUs for large-scale training. The source does not attribute that claim to any independent verification.
Why it matters
Meta operates Nvidia GPUs, AMD GPUs and its own MTIA chips partly to avoid depending on one vendor. This paper is the clearest evidence yet that doing so requires rewriting models and tooling, not just drivers. Anyone evaluating AMD for training workloads now has a public playbook. Upstreamed fixes in PyTorch, Triton and ROCm lower the cost of following it. That erodes one argument for staying locked to Nvidia's software stack by default.
Primary source
arXiv cs.LG — arxiv.orgWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.