AMD has released 'ROCm 10,' which enhances the AI processing capabilities of AMD GPUs, and has also introduced 'ROCm.AI,' which provides skills and other resources for AI development.
AMD is deploying ROCm as a GPU computing software stack to compete with NVIDIA's CUDA. ROCm 10 was released on August 27, 2026. Its key features include the AI-driven development platform 'ROCm.AI,' as well as optimizations to speed up image and video generation in ComfyUI.
ROCm 10.0: A Decade of Open Compute, Built for the Age of Agentic AI — ROCm Blogs
AMD ROCm 10: Bringing ROCm.AI's AI-Native Developer Experiences to AMD Platforms - AMD Newsroom
https://newsroom.amd.com/news/rocm-10-software-ai-native-developer-experiences/
AMD ROCm.AI | ROCm™ Software
https://www.amd.com/en/products/software/rocm/rocm-ai.html
In the fields of AI learning and AI inference, implementations based on NVIDIA's CUDA are dominant, and currently, AMD GPUs, even those with equivalent specifications to NVIDIA GPUs, lag behind in AI processing power. To overcome this situation, AMD is focusing on developing ROCm, a system that competes with CUDA. ROCm 10 is a major version following ROCm 7, which was released in September 2025, and was named 'ROCm 10' by skipping '8' and '9'.
One of the key features being touted as a highlight of ROCm 10 is ROCm.AI. ROCm.AI is an AI-driven development platform comprised of AI-related technologies such as 'AMD Skills,' 'Hyperloom,' and 'ROCm CLI,' and is said to significantly reduce the manual work required by developers.
AMD Skills are skills usable with AI agents such as Claude Code and Codex, allowing you to provide AI agents with 'verified ROCm knowledge' from AMD. It covers not only AI-dedicated GPUs like the MI300X, but also CPU-related knowledge such as 'quantization techniques for performing AI inference on the CPU.' You can find more information about AMD Skills at the following link.
GitHub - amd/skills: Official AMD catalog of AI agent skills. Empower your AI agents with AMD's optimized SW stack. · GitHub
https://github.com/amd/skills
Hyperloom is an agent system for building end-to-end inference workloads that completely eliminate human intervention. It enables complex tasks such as software optimization by having AI loop through operations such as profiling, analyzing results, considering changes, applying changes, and validating correctness during the workload. AMD claims that it can reduce tasks that would take weeks manually to just a few hours. Hyperloom documentation is available at the following link.
Hyperloom documentation — Hyperloom 1.0.0 documentation
https://rocm.docs.amd.com/projects/hyperloom/en/latest/
Furthermore, ROCm 10 includes inference optimizations specifically for ComfyUI. This has resulted in improved generation speeds for video generation AI 'Wan2.2' and image generation AIs such as 'FLUX.2 KLEIN,' 'Stable Diffusion 3.5 Medium,' 'Stable Diffusion 2.1,' and 'Stable Diffusion XL Base.'
AMD plans to release minor updates to ROCm 10 approximately every six weeks.
Related Posts:
