AMD released ROCm 10 on 27 August at 15:09 UTC and made ROCm.AI — Hyperloom, AMD Skills and the ROCm CLI — generally available. The headline claim: "a system configured with ROCm.AI delivers an average 3.3x inference improvement and 2.4x training improvement over ROCm 7 on the same hardware."

What the footnote says

Footnote MI350-81, on the same page: "Testing by AMD Performance Labs as of July 7, 2026 … an AMD Instinct MI355x 8x GPU platform and AMD ROCm 7.0 software vs a similarly configured system using a preview version of AMD ROCm.ai (ROCm 7.2.2 with optimizations such as Optimized Kernels, Parallelism and Scheduling) running GLM-5, Kimi-K2.5, and DeepSeekk-R1-0528 models. Stated performance uplift is expressed as a combined average TPS over across the (3) models tested." The training footnote, MI350-82, carries the same July date against Megatron-LM on DeepSeek-V2-Lite, DeepSeek-V3-16B and Qwen3-30B-A3B.

What actually shipped

Hyperloom is described as an "autonomous agentic system for optimizing end-to-end inference workloads," supporting vLLM and SGLang and targeting HIP, Triton and FlyDSL. AMD Skills ship into the Claude Code, Cursor and Codex marketplaces plus an open GitHub catalogue. That is the genuinely novel part: AMD is using agents to tune kernels rather than waiting on compiler work.

What the common framing gets wrong

Every headline will read "ROCm 10 is 3.3x faster." AMD's own footnotes contradict that three ways. The 3.3x was not measured on ROCm 10 — it is ROCm 7.0 versus a preview build of ROCm 7.2.2, benchmarked seven weeks before this release. It is not 3.3x against Nvidia: the baseline is AMD's own year-old software on the same MI355X hardware, so the number measures how much performance AMD had been leaving on the table, not a competitive position. And it is a combined average across exactly three models, not a general speedup. One more caveat worth naming: ROCm CLI, one of the three pillars of the "generally available" ROCm.AI, is labelled a Technology Preview and, in AMD's words, does not officially support ROCm 10 yet — "beginning with ROCm 7.13 software, with ROCm 10 official support coming soon."