AMD has released ROCm 10, marking 10 years since the first version of its open source compute platform and adding a broader AI focused development environment for Instinct accelerators, Radeon GPUs and Ryzen integrated graphics.
The update introduces ROCm.AI, a new developer layer built around ROCm CLI, AMD Skills and Hyperloom. AMD says systems using ROCm.AI can deliver an average 3.3x improvement in inference performance and a 2.4x gain in training performance compared with ROCm 7 under its internal testing conditions.
| ROCm 10 feature | Main role |
|---|---|
| ROCm.AI | AI focused developer workflow |
| ROCm CLI | Install, serve, validate and troubleshoot workloads |
| AMD Skills | ROCm knowledge for AI coding assistants |
| Hyperloom | Automated workload optimization |
| Supported hardware | Instinct, Radeon and Ryzen integrated graphics |
| Platforms | Windows and Linux |
| Reported inference uplift | 3.3x versus ROCm 7 |
| Reported training uplift | 2.4x versus ROCm 7 |
ROCm.AI changes the developer workflow
ROCm 10 is designed to move beyond providing low level libraries and compute tools by improving the process of setting up, running and optimizing AI workloads.
ROCm.AI sits at the center of that effort.
Its first component is ROCm CLI, a unified command line tool that can install, validate, serve, update and troubleshoot AI environments. AMD is releasing the CLI as a technology preview, so its interface and capabilities may continue to change.
Commands include the ability to launch a model for inference and diagnose issues related to drivers or environment configuration. The tool also supports air gapped systems by allowing dependencies to be packaged for offline installation.
AMD Skills brings ROCm guidance into coding assistants
AMD Skills is intended to make ROCm specific knowledge available directly inside AI coding tools.
The system uses the Agent Skills format and supports assistants including Claude, Cursor and Codex. AMD maintains a catalog of validated skills that can be installed into the directories already used by these development tools.
The goal is to reduce the amount of manual searching and configuration needed when developers are working with AMD hardware and ROCm software.
Instead of separately consulting documentation for common deployment or optimization tasks, developers can access AMD prepared guidance from within the coding assistant they are already using.
Hyperloom automates AI workload optimization
Hyperloom is the most automated part of the new ROCm.AI toolset.
AMD describes it as an open source agentic system capable of handling an entire inference optimization cycle. It can establish a performance baseline, profile a workload, identify bottlenecks, plan changes, modify serving configurations or GPU kernels and then validate whether those changes actually improved performance.
The cycle can then repeat without requiring an engineer to manually control every stage.
AMD claims this approach can reduce optimization work that previously took weeks to a matter of hours. That claim should be treated as AMD's own performance assessment, since results will depend on the model, hardware and workload being optimized.
AMD reports large inference and training gains
AMD says ROCm.AI delivered around 3.3x higher inference throughput and 2.4x higher training performance compared with ROCm 7 in its testing.
The inference comparison used an eight GPU Instinct MI355X platform and workloads including GLM 5, Kimi K2.5 and DeepSeek R1 0528.
The comparison also involved a preview ROCm.AI configuration with optimizations related to kernels, parallelism and scheduling, so the result should not be interpreted as a universal 3.3x uplift simply from installing ROCm 10 on any existing system.
The broader significance of ROCm 10 is its focus on reducing setup and optimization work across AMD's AI hardware stack.
Instinct accelerators, Radeon graphics cards and Ryzen integrated GPUs are now covered under the same ROCm 10 framework on both Windows and Linux. With ROCm CLI simplifying deployment, AMD Skills bringing validated guidance into AI assistants and Hyperloom automating optimization, AMD is trying to make its hardware easier to use for AI development as well as faster once workloads are running.
Discussion (0)
Be the first to comment.