AMD ROCm 10: ROCm.AI Brings Skills, Hyperloom, CLI

AMD ROCm 10 ships with ROCm.AI generally available

ROCm is the software AMD provides for running AI training, inference and scientific computing on its GPUs. It does the job CUDA does on NVIDIA hardware, and its compilers, runtimes and libraries are open source.

AMD released ROCm 10 on August 27, 2026, announcing it in its newsroom, and made ROCm.AI generally available at the same time. ROCm.AI had only been previewed at Advancing AI 2026, and from this release anyone can install it.

One stack for Instinct, Radeon, and Ryzen GPUs

TheRock, AMD's open-source build and release system, produced every part of ROCm 10, the first major release to be built that way. TheRock entered production with ROCm 7.14. One pipeline builds the libraries and framework packages and checks the whole stack before anything is released.

With a single pipeline behind it, the same ROCm now runs on Instinct data center accelerators, Radeon graphics cards and Ryzen integrated graphics, on Windows and Linux alike. A feature added for one GPU family or operating system gets built and tested for the rest in the same release. AMD's pitch is that what you prototype on a laptop carries over to a data center rack.

Version 10 lands a decade after ROCm 1.0 came out in April 2016, and ROCm Blogs says skipping 8 and 9 was on purpose. Minor releases will keep coming about every six weeks.

The Windows HIP SDK gives way to one Core SDK

The Windows HIP SDK is gone in ROCm 10. ROCm Core SDK replaces it, which puts Windows on the same SDK definition and release schedule as Linux. Windows developers used to get a separate package on its own timeline, and now both systems share one SDK and one release schedule.

Packaging is the weak spot for now. Windows builds come as a tarball that you unpack and point your build settings at, and AMD says native Windows installers with managed updates will arrive later in 2026.

ROCm.AI in three parts: Skills, Hyperloom, and a CLI

Earlier ROCm releases mostly improved the libraries. ROCm.AI goes after the work around them, from installing ROCm to running models and speeding them up. Two of its three parts are generally available, and the CLI is still a Technology Preview.

AMD Skills bring AMD know-how to Claude Code, Codex, Cursor

AMD Skills hand AI coding agents ROCm knowledge that AMD has checked, so the agents can use it in the middle of a task. They use the Agent Skills format and drop into folders those tools already read, such as ~/.claude/skills/ for Claude Code.

The skills install from the Claude Code, Codex and Cursor marketplaces or from the open amd/skills catalog on GitHub. That catalog has three areas: local AI on client PCs, cross-stack jobs such as diagnostics and optimization, and server work on Instinct GPUs and EPYC CPUs. One skill in it, serving-llms-on-instinct, takes an agent through standing up a vLLM server with an OpenAI-compatible API on Instinct hardware.

Every skill that ships has passed structural and behavioral tests, AMD says, to keep workflows consistent and reliable. An agent setting up an environment or serving a model therefore starts from steps AMD has already checked. For Claude itself, see What Is Claude AI? Free Plan, Models, vs ChatGPT.

Hyperloom tunes vLLM and SGLang inference on Instinct

ROCm Hyperloom is an open-source agent that keeps improving inference speed on its own, round after round, without a person directing each step. Its targets are vLLM and SGLang, two open-source servers for running large language models.

In every cycle Hyperloom profiles the workload, locates the bottleneck, writes candidate fixes, benchmarks them and checks both speed and output correctness. Its changes can go into GPU kernels written in HIP, Triton or FlyDSL, or into the host code that calls them. At the end it returns a report of the proposed edits with their measured or expected gains.

AMD says weeks of manual tuning shrink to hours. The catch is hardware: AMD lists Hyperloom only for Instinct MI300X, MI325X and MI355X and says nothing about Radeon or Ryzen. It installs with a single pip command and runs either alone or through AMD Skills.

The CLI is a Technology Preview for Windows and Linux

The ROCm CLI puts installing, troubleshooting and serving models into one command-line tool, and in ROCm 10 it carries the Technology Preview label. AMD distributes it as a prebuilt binary for Windows and Linux that works without ROCm already installed.

rocm serve <model> starts a model for inference, and rocm examine looks for environment and driver problems. Several ROCm versions can sit side by side, and you can switch between them or roll back. The bundled ROCm Console, once called dash, shows GPU utilization and, on supported Instinct systems, HBM memory use, power draw and tokens per watt.

The CLI is early access, and AMD expects its interface to keep changing. It works with ROCm releases from 7.13 on, but AMD lists official ROCm 10 support as coming soon, so read the release notes before relying on it with ROCm 10.

Reading the 3.3x claim and the gap with CUDA

AMD's headline claim is that ROCm.AI makes inference 3.3x faster and training 2.4x faster on average. AMD measured both figures in its own lab, and the footnotes set conditions few readers' setups will match.

The 3.3x and 2.4x figures come from eight MI355X GPUs

AMD Performance Labs ran both tests on July 7, 2026, on one system with eight Instinct MI355X GPUs. The inference figure averages tokens per second over three models, and the training figure averages three other models trained with the Megatron-LM framework. The footnotes give these conditions.

The comparison row matters most. The baseline is ROCm 7.0, and the faster side is a preview of ROCm.AI on ROCm 7.2.2 with added kernel, parallelism and scheduling work, not ROCm 10 as it ships.

The serving software also differs between the two sides of the inference test: Kimi-K2.5 goes from vLLM 0.16.0 to 0.22.0, GLM-5 moves from SGLang to ATOM, and only DeepSeek-R1 keeps the same SGLang version. AMD adds that results vary with configuration, software and drivers. Installing ROCm 10 will not by itself make your GPU 3.3x faster.

Radeon and Ryzen AI Max get ComfyUI tuning and Unsloth

On Radeon and Ryzen, what ROCm 10 offers is tuned image and video models, plus local fine-tuning on Ryzen AI Max only. The 3.3x figure comes from MI355X and Hyperloom is listed for Instinct only, so neither carries over directly.

ComfyUI workflows for image and video models such as Wan2.2, FLUX.2 KLEIN and Stable Diffusion 3.5 Medium have been tuned for supported Radeon and Ryzen hardware. Ryzen AI Max machines, with their large pool of unified memory, get Unsloth support for LoRA and QLoRA fine-tuning in ROCm 10, which lets you adapt a model locally instead of renting cloud GPUs.

Radeon and Ryzen also now have production support for vLLM and SGLang, with validated containers on Docker Hub. Running models on your own machine is covered in Local LLM Coding Now Hits 75% of Cloud AI.

ROCm 10's pitch against CUDA is developer tooling, such as coding-agent skills and a single CLI, on top of the usual library updates. Gaps remain: the CLI is a preview, and native Windows installers will not arrive until later in 2026. AMD published no ROCm-versus-CUDA numbers, so anyone thinking of switching should benchmark their own workload on both.

Summary: ROCm 10 unifies AMD GPUs, but the 3.3x is a lab figure

ROCm 10 celebrates a decade of ROCm by skipping to version 10 and putting Instinct, Radeon and Ryzen on one stack. ROCm.AI became generally available the same day, bringing AMD Skills for Claude Code, Codex and Cursor, Hyperloom for automated inference tuning, and a preview ROCm CLI for setup and diagnostics.

The 3.3x inference figure is AMD's own measurement on eight MI355X GPUs, and Hyperloom is listed for Instinct only. On Radeon or Ryzen hardware, start with the ComfyUI tuning, or with Unsloth fine-tuning on Ryzen AI Max.