AMD Launches ROCm 10 with AI-Native Developer Tools for AMD Platforms

AMD has officially launched ROCm 10, marking a decade of its open software stack while making ROCm.AI generally available. The latest release introduces AI-driven development and optimization tools designed to help developers build, deploy and improve AI workloads on AMD hardware.

ROCm.AI brings together three key technologies: ROCm Hyperloom, AMD Skills and ROCm CLI. Together, they provide AMD-specific expertise, simplified workflows and agentic optimization directly within development environments.

According to AMD, AI-driven optimization of kernels, memory management and scheduling can deliver an average 3.3x improvement in inference performance and 2.4x improvement in training performance compared with ROCm 7 on the same hardware.

Autonomous AI Optimization with ROCm Hyperloom

A major part of ROCm.AI is ROCm Hyperloom, an autonomous agentic system designed to optimize end-to-end AI inference workloads.

Hyperloom can profile workloads, identify bottlenecks, test optimization strategies, implement changes, benchmark results and validate performance and correctness. With ROCm 10, it expands support across AMD Instinct GPUs, including workloads using vLLM and SGLang.

Developers can also target optimizations across HIP, Triton and FlyDSL, with reports outlining proposed changes and expected or measured performance improvements.

AMD Skills Brings Expertise into Coding Tools

AMD Skills brings curated AMD knowledge and validated workflows into AI coding assistants such as Claude Code, Cursor and Codex.

The expanded Skills catalog covers client-native workflows for local AI applications, cross-stack workflows for diagnostics and optimization, and server-native workflows targeting AMD Instinct GPUs and AMD EPYC processors.

AMD Skills are available through the respective coding-agent marketplaces as well as an open catalog on GitHub.

ROCm CLI Simplifies AI Workflows

ROCm 10 also introduces the ROCm CLI, currently available as a Technology Preview. The unified command-line interface is designed to simplify setting up, managing and running AI workloads on AMD hardware.

Developers can use it to inspect systems, manage ROCm environments, serve models, run diagnostics and control runtimes. It supports both Windows and Linux as a prebuilt binary and does not require an existing ROCm installation.

The included ROCm Console provides real-time monitoring of runtime health, model serving, GPU utilization and benchmark information, including HBM usage, power consumption and tokens-per-watt metrics on supported AMD Instinct systems.

A Broader ROCm 10 Upgrade

Beyond ROCm.AI, ROCm 10 also expands the wider ROCm software ecosystem with a more modular ROCm Core SDK and updates across libraries, compilers, frameworks, tools, model support and hardware platforms.

With ROCm.AI now generally available, AMD is positioning ROCm 10 as a major step toward more AI-native development workflows, helping developers make better use of AMD GPUs and processors while reducing the complexity of performance optimization.

The release marks an important milestone for AMD’s ROCm platform as AI development increasingly shifts toward automated, agent-assisted optimization.