AMD Releases ROCm 10 and Makes ROCm.AI Generally Available

ROCm 10 ships with ROCm.AI generally available

AMD released ROCm 10 on August 27, 2026 and made ROCm.AI generally available. ROCm.AI is an AI-native developer layer built around three core experiences: ROCm Hyperloom, AMD Skills and the ROCm CLI.

The release distinction matters. ROCm.AI as a broader platform is generally available with ROCm 10, while the ROCm CLI itself is still described by AMD as a Technology Preview. AMD also says official ROCm 10 support in the version-agnostic CLI is coming soon, even though the CLI can already manage environments beginning with ROCm 7.13.

Hyperloom is an autonomous inference-optimization agent

ROCm Hyperloom profiles an inference workload, identifies bottlenecks, explores optimization choices, implements targeted changes, benchmarks the result, and validates performance and correctness. AMD says ROCm 10 expands Hyperloom support across AMD Instinct GPUs and adds support for vLLM and SGLang, with optimization targets spanning HIP, Triton and FlyDSL.

That makes Hyperloom materially different from a passive profiler. It is intended to act as an autonomous optimization system across host code and GPU kernels, while producing reports about proposed changes and their measured or expected impact.

AMD Skills puts ROCm-specific workflows into coding agents

AMD Skills packages curated AMD knowledge and validated workflows for supported AI coding agents. AMD says the skills are now available through Claude Code, Codex and Cursor marketplaces, as well as an open catalog on GitHub.

The expanded catalog covers local/client AI workflows, cross-stack diagnostics and optimization, and server workflows for AMD Instinct GPUs and AMD EPYC processors. AMD says shipped skills pass structural and behavioral tests before release.

For AI developers, this is an important developer-platform change: hardware-specific guidance and optimization procedures can be invoked from the coding agent already being used rather than requiring developers to manually translate every ROCm diagnostic or tuning task into agent instructions.

The ROCm CLI unifies environment and model operations

The ROCm CLI is intended to provide one interface for installing and managing ROCm environments, running diagnostics, controlling runtimes and serving models. AMD says it supports Windows and Linux as a prebuilt binary and does not require an existing ROCm installation.

The CLI can manage multiple side-by-side runtimes, activation and rollback, model serving and engine management. AMD also includes ROCm Console for runtime health, model-serving activity, GPU utilization and supported telemetry such as HBM use, power and tokens per watt.

Because the CLI remains a Technology Preview, it should not be presented as a fully mature generally available component simply because the broader ROCm.AI experience is GA.

AMD reports large performance uplifts, with important test caveats

AMD reports an average 3.3× inference improvement and 2.4× training improvement for systems configured with ROCm.AI optimizations compared with ROCm 7 on the same hardware. The figures come from AMD Performance Labs and combine results across specific models, software versions and AMD Instinct MI355X configurations.

Those numbers are vendor-reported test results, not universal ROCm 10 speedups. AMD's own notes state that performance varies with configuration, software, drivers, optimization choices and serving or training stacks.

ROCm 10 also changes the underlying software stack

Beyond the agentic developer layer, ROCm 10 introduces a more modular ROCm Core SDK and updates libraries, compilers, frameworks, tools, model support, performance and hardware-platform coverage.

The strategic significance is that AMD is pairing GPU software-stack improvements with agentic tooling that can increasingly diagnose, tune and operate AI workloads. That puts software automation and developer experience alongside raw accelerator performance as part of AMD's effort to compete for production AI workloads.