AMD released ROCm 10 on August 27, 2026, and made ROCm.AI generally available as an additional set of developer experiences for its AI software stack. ROCm.AI combines ROCm Hyperloom for workload optimisation, AMD Skills for hardware-specific coding guidance and the ROCm CLI for managing environments and AI workloads.
The release marks the tenth anniversary of AMD’s ROCm software stack. AMD also describes ROCm 10 as a broader update to the underlying platform, including a more modular Core SDK and changes across libraries, compilers, frameworks, tools, model support and hardware platforms.
Contents
Three components for different development tasks
ROCm is AMD’s software stack for working with its computing hardware. An AI stack normally includes compilers, libraries, frameworks, runtimes, developer tools and hardware-specific optimisation. ROCm.AI addresses several of those layers without replacing the broader ROCm SDK.
The three ROCm.AI components have different roles:
- ROCm Hyperloom is designed to find and apply performance optimisations to inference workloads.
- AMD Skills provides AMD-specific instructions and validated workflows inside supported AI coding agents.
- The ROCm CLI provides a command-line interface for setting up, operating and monitoring AI workloads on AMD hardware.
This division is significant because deploying an AI model involves more than selecting a GPU. Developers may need to identify bottlenecks, adjust GPU kernels or parallelism, manage compatible runtime versions, deploy a model-serving engine and monitor performance during operation.
What the tools do
Hyperloom profiles and changes workloads
AMD describes Hyperloom as an autonomous agentic system. In this context, an agentic system is software that can perform a sequence of tasks toward a goal rather than only returning a single recommendation.
According to AMD, Hyperloom profiles an inference workload, identifies bottlenecks, explores optimisation options, implements changes, benchmarks the result and validates both performance and correctness. Its optimisation scope can include host code and GPU kernels rather than being limited to one library or kernel.
With ROCm 10, AMD says Hyperloom expands support across AMD Instinct GPUs and supports the vLLM and SGLang serving frameworks. Developers can target optimisation work across HIP, Triton and FlyDSL, which are different programming and development environments used to express GPU-accelerated computation.
The practical objective is to reduce the manual work involved in adapting model-serving code to AMD hardware. The announcement does not, however, quantify how much time or engineering effort Hyperloom saves compared with conventional optimisation work.
AMD Skills brings platform guidance into coding agents
AMD Skills is intended to place AMD-specific guidance inside supported AI coding agents, including Claude Code, Cursor and Codex. AMD says its catalogue includes client-native, cross-stack and server-native workflows, including workflows for AMD Instinct GPUs and AMD EPYC processors.
The released skills are available through the Claude Code, Codex and Cursor marketplaces and through an open catalogue on GitHub, according to AMD. The company says each shipped skill undergoes structural and behavioural testing before release.
These skills are not described as a replacement for general-purpose coding agents. Instead, they are intended to give those agents information and workflows tailored to AMD’s hardware and software stack. Their usefulness will depend on the quality of the guidance, the supported versions and how well the resulting code performs on a particular workload.
The ROCm CLI manages environments and serving
The ROCm CLI is a unified command-line interface for setting up, managing and operating AI workloads on AMD hardware. AMD says it is available as a prebuilt binary for Windows and Linux and does not require an existing ROCm installation.
The CLI supports managed ROCm environments, side-by-side runtimes, runtime activation and rollback, model serving and engine management. This makes it broader than an installation-only utility: developers can use it to control runtime versions and serving workflows after the initial setup.
The included ROCm Console provides views of system status, workload activity, runtime health, model serving, GPU utilisation and benchmark telemetry. On supported AMD Instinct systems, it can expose high-bandwidth memory usage, power consumption and tokens-per-watt metrics.
AMD says the same CLI workflows can be used manually, through an AI coding agent or in continuous-integration environments. However, the CLI remains a Technology Preview component. It supports workflows beginning with ROCm 7.13 software, while official ROCm 10 support is listed as coming soon. That means the general availability of ROCm.AI does not imply that every component has the same maturity or support status.
AMD’s performance claims
AMD says ROCm.AI delivered an average 3.3x improvement in inference and 2.4x improvement in training compared with ROCm 7 on the same hardware. Inference means running a trained model to generate outputs; training means adjusting a model’s parameters using data. The reported results were measured in tokens per second, a throughput metric commonly used for language-model workloads.
The figures come from separate tests by AMD Performance Labs on July 7, 2026. The inference comparison used an eight-AMD-Instinct-MI355X system and the GLM-5, Kimi-K2.5 and DeepSeek-R1-0528 models. The training comparison used an eight-GPU AMD Instinct MI355X system with Megatron-LM and DeepSeek-V2-Lite, DeepSeek-V3-16B and Qwen3-30B-A3B.
The comparisons used the same specified hardware while changing the software and optimisation stack. AMD describes the inference test as a comparison between ROCm 7.0 and a ROCm.AI preview based on ROCm 7.2.2. The training comparison used ROCm 7.0 against a ROCm.AI preview using ROCm 7.2.1 and Primus v26.3.
Keeping the hardware constant can help isolate some software-related differences. It does not make the result universal: throughput remains sensitive to the model, software versions, drivers, configuration, serving or training framework and measurement method.