AMD MI350X Matches B200 in Meta Training Test

Meta engineers say AMD's Instinct MI350X graphics processors now train the company's production recommendation models at roughly the same speed as Nvidia's B200. In a paper posted to arXiv on Oct. 15, 2026, the team reported a geometric-mean throughput of 101.2% of its B200 baseline across 10 production models, with tens of thousands of MI350X chips already running training jobs at Meta. AMD publicized the work on Oct. 27 alongside an expanded partnership with Meta.

Key takeaways

• Across 10 production recommendation models, MI350X averaged 101.2% of B200 training throughput; it was faster on four models and slower on six, with results ranging from 82% to 130%.
• AI agents helped port CUDA-era code, with a success rate above 80% on dozens of production Triton kernels, according to the paper.
• The comparison used a 1,000-watt MI350X against a B200 capped at 750 watts, and it covers recommendation models, not frontier language-model training.

AMD MI350X vs. Nvidia B200: Meta's benchmark results

The paper, "Architecting for Hardware Diversity: Meta's End-to-End Co-Design for High-Performance AI Training on AMD GPUs," lists 58 authors and calls itself the first hyperscaler report on putting AMD GPUs into large-scale training production. Its Table III compares end-to-end training throughput on matched GPU counts, from eight to 128 chips.

The authors credit three changes: redesigning models and training recipes for the hardware instead of porting them directly, an analytical model that picks matrix-math kernels without warmup runs, and agent-driven porting. Together, they say, these deliver up to 1.5 times the throughput of a straight port.

How AI agents helped port CUDA code to ROCm

Meta's agents compile code, read error messages and apply fixes "until the build succeeds," the paper says, and can write shims where a CUDA library has no ROCm equivalent. On dozens of production Triton kernels, the workflow succeeded more than 80% of the time.

In one case, an agent rewrote a jagged flash attention backward kernel to remove atomic operations, cutting its run time from 32 milliseconds to 14 milliseconds. MI350X went from 2x slower than B200 on that kernel to 15% to 20% faster, and the model's end-to-end training throughput rose 25%.

The caveats: power, memory and model type

The matchup is not like-for-like on power. Meta runs MI350X at 1,000 watts and B200 at 750 watts on the same air-cooled platform, both below vendor limits; Nvidia's general HGX B200 is rated at 1,000 watts. MI350X also carries 288 GB of HBM3e memory, about 1.6 times the B200's 180 GB, which Meta uses for larger batches and embedding tables.

Gaps remain: matrix-math speed is limited partly by emulated TF32 support, the paper says.

Context: AMD and Meta expand their partnership

In an Oct. 27 press release, AMD said it will invest $1 billion in Meta's equity offering and that initial Helios rack-scale systems are installed and being tested at Meta. Meta was an early customer for the MI450 architecture and will deploy a custom Instinct GPU, AMD said. "AMD has been a key partner in building out the compute that's running our frontier models," Meta CEO Mark Zuckerberg said in the release.

Models compatible with AMD GPUs now make up more than 60% of Meta's production model types, weighted by resource use, the paper says. Meta also runs Nvidia GPUs and its own MTIA chips to cut costs and avoid vendor lock-in.

What's next

Meta says its changes have been upstreamed to PyTorch, Triton and ROCm, so other ROCm users could benefit. AMD said the joint work informs its future MI500 series. Neither company gave a timeline for broader Helios deployment beyond "the near future."

FAQ

Did AMD beat Nvidia in Meta's tests?

On average the two were roughly even, at 101.2%. MI350X won on four of 10 models and lost on six, and it drew more power per chip in Meta's setup.

Does this apply to large language model training?

The benchmark table covers recommendation models. The paper says Meta has also trained LLMs on clusters of more than 2,000 MI350X GPUs but does not give a B200 comparison for them.

References

• Architecting for Hardware Diversity: Meta's End-to-End Co-Design for High-Performance AI Training on AMD GPUs — arXiv (Oct. 15, 2026)
• AMD and Meta Advance Strategic Partnership to Power Meta's Superintelligence Infrastructure — AMD (Oct. 27, 2026)
• AMD Beat Nvidia on 4 of 10 Meta Training Models, Paper Shows — FourWeekMBA (Oct. 30, 2026)