AMD announced ROCm.AI at its Advancing AI event in San Francisco on July 24, introducing a platform aimed at improving AI model inference performance on Instinct GPUs.

For years, AMD GPUs have been perceived as less capable than competing products because they do not support Nvidia’s CUDA stack, but recent AI frameworks such as PyTorch and JAX allow code to run across hardware without direct CUDA interaction.

The ROCm.AI platform integrates with AI code assistants and provides documentation, debugging tools and an automated optimization system called Hyperloom. Hyperloom can launch a Docker container, run benchmark baselines, profile workloads and adjust configurations or generate custom CPU kernels on demand.

In testing on AMD Helios racks, AMD reported a 38 % increase in inference speed compared with baseline measurements. Anush Elangovan, AMD corporate vice president of AI software and solutions, said the system is intended to let users easily consume, debug and deploy optimized models.

AMD plans to collaborate with model developers including OpenAI and Anthropic to ensure native support for AMD programming, and ROCm.AI will be offered as a plug‑in for coding assistants such as Anthropic Claude Code, OpenAI Codex, Google Antigravity and Cursor.