AMD's ROCm.ai makes coding agents tune your AMD GPUs, 3.3x faster
Announced at Advancing AI, shipping in August: a ROCm CLI, AMD Skills that brief Claude, Cursor and Codex, and open-source Hyperloom you can clone now.
Copy markdown
Three parts: a CLI, agent Skills, and an optimizer
ROCm.ai bundles a unified ROCm CLI, AMD Skills that turn Claude, Cursor, Codex and Gemini into ROCm “superusers,” and Hyperloom, an agentic system that profiles and tunes workloads. AMD's Anush Elangovan frames it as agents that don't just answer questions but “profile, debug and drive workloads toward peak performance.”
3.3x inference, 2.4x training — vs ROCm 7
On identical hardware, AMD claims an average 3.3x inference and 2.4x training uplift over ROCm 7. ROCm 7.14 is out now; the ROCm.ai experience lands with the August release. Treat the multipliers as vendor numbers until third parties benchmark them.
Hyperloom is MIT-licensed and on GitHub today
You don't have to wait for August. Hyperloom is already public under the MIT license: it profiles with TraceLens, tunes hot kernels with GEAK, and searches optimizations with Arbor in a Think-Decide-Implement-Benchmark loop. Clone it and point it at an LLM workload on an AMD GPU.
The real target: CUDA's tooling moat
Builders default to NVIDIA because of CUDA's developer experience, not only its silicon. Baking AMD-specific knowledge into the agents you already run, plus one CLI to install, serve and debug, aims squarely at that switching cost — the first credible dent in it.
Local dev: Ryzen AI Halo and the usual frameworks
AMD also pushed the Ryzen AI Halo platform for local AI development, with more Halo systems due later this year, and says PyTorch, vLLM, SGLang and Hugging Face already run on its new MI455X GPUs.