ROCm 10.0 Marks a Decade of Open Compute, Targeting Agentic AI Development

The release of ROCm 10.0 represents a major milestone for AMD’s open-source GPU compute stack. Since first shipping ROCm 1.0 in April 2016, the platform has grown significantly, evolving into a system capable of training and executing advanced AI models across numerous industries. The 10.0 version marks this decade-long journey with substantial architectural upgrades, most notably the introduction of ROCm.AI, an AI-focused developer experience designed for modern autonomous workflows.

Core Architectural Enhancements and Distribution

The ROCm Core SDK 10.0 is the first major version increment since the 7.x series. This entire release is built upon TheRock, an automated open-source build and release system that achieved production status with ROCm 7.14. This single, unified pipeline ensures that all components, including libraries, primitives, and framework wheels, are staged and validated across the entire stack before release. Developers can anticipate that minor updates to ROCm 10.0 will follow at an approximate interval of six weeks.

This standardized approach allows the platform to operate seamlessly across AMD’s entire portfolio—including Instinct accelerators, Radeon graphics cards, and Ryzen integrated graphics—on both Windows and Linux. Critically, the development and validation process is unified: any capability proven on a laptop is instantly available for use in a data center environment, eliminating the need for separate porting efforts.

AMD is also consolidating its GPU software distribution under a single repository at repo.amd.com. Previously, users often had to gather necessary components—such as ROCm packages, the amdgpu driver, and specialized data center tools—from multiple, disparate sources, each with unique instructions. The new repository structure centralizes ROCm packages, the amdgpu driver, and public GPU tools, providing a consistent layout and configuration method. This consolidation allows a single machine to manage multiple ROCm versions and architectures simultaneously, giving teams flexibility over the release cycle.

ROCm.AI: A Unified AI Developer Workflow

The most significant feature of ROCm 10.0 is ROCm.AI, which is AMD’s new, AI-native developer experience. Unlike previous releases that focused primarily on providing improved building blocks and libraries, ROCm.AI focuses on streamlining the entire AI development workflow: from setup and validation to deployment and optimization on AMD hardware, significantly reducing manual effort.

ROCm.AI is comprised of three integrated components:

• ROCm CLI: This is a singular command-line interface designed to manage AI workloads on AMD platforms. It replaces a collection of disparate scripts and manual setup procedures. The CLI supports functions such as running inference for a model using rocm serve <model>, diagnosing environment issues with rocm examine, and supports air-gapped environments by managing dependencies in a self-contained bundle. This tool is available in a tech preview state in ROCm 10.0.
• AMD Skills: This feature integrates AMD-validated ROCm knowledge directly into popular AI coding assistants, such as Claude, Cursor, and Codex, utilizing the Agent Skills format. The official catalog is maintained in the amd/skills GitHub repository. This catalog supports both GPU and CPU workflows. For the GPU side, rocm-doctor assists diagnostics, and serving-llms-on-instinct walks an agent through setting up a vLLM OpenAI-compatible endpoint on MI300X, MI325X, MI350X, or MI355X. For CPU environments, the catalog extends this capability to EPYC processors using ZenDNN and zentorch, ensuring a consistent, skill-driven workflow regardless of the target hardware.
• Hyperloom: This is an advanced, open-source, agentic system that aims to automate the entire process of optimizing inference workloads end-to-end. Instead of requiring manual tuning cycles, Hyperloom automates the loop of Profiling → Analyzing → Planning → Optimizing → Validating. AMD reports that this system can reduce optimization work that previously took weeks of human effort down to mere hours, while exploring a vastly larger solution space. The system orchestrates several components, including TraceLens-Agent, Magpie, IntelliKit’s conversational profiler, GEAK’s autonomous multi-agent kernel optimization, and Arbor’s search mechanism. Hyperloom is currently operational on MI300X, MI325X, and MI355X and can be initiated via a simple pip command.

Performance Libraries and Framework Support

ROCm 10.0 provides production-ready containers and support for vLLM and SGLang, enabling teams to run Large Language Model (LLM) inference on AMD Instinct, Radeon, and Ryzen hardware without needing to build the stack from source. These validated containers and Python wheels are available on Docker Hub, built through TheRock’s CI pipeline.

Additional performance enhancements include:

• Local LLM Tuning: Support for Unsloth on Ryzen™ AI MAX platforms allows for fast and memory-efficient local fine-tuning of LLMs. Through techniques like LoRA and QLoRA, developers can perform model tuning locally, leveraging the large unified memory available on Ryzen™ AI MAX, thereby enhancing data privacy and accessibility.
• Generative AI Optimization: ComfyUI has received expanded optimizations for leading image and video generation models, including Wan2.2, FLUX.2 KLEIN, Stable Diffusion 3.5 Medium, Stable Diffusion 2.1, and Stable Diffusion XL Base. AMD has also published guidance on selecting optimal attention algorithms and backends for ComfyUI, helping users achieve high performance with minimal configuration changes.
• Communication Libraries: Significant upgrades were made to communication libraries to narrow the gap with competitors.
  • RCCL (ROCm Collective Communications Library): The upstream merge from NCCL advanced to version 2.30.4. Key additions include symmetric memory support for tightly-coupled multi-GPU peer-to-peer communication, a GPU-initiated networking (GIN) device API (allowing GPUs to initiate network transfers directly), one-sided host APIs, and new Pythonic APIs for direct use within Python code.
  • rocSHMEM: This library continues to close the gap with NVSHMEM 3.6.5. This release introduces host AMO and context APIs with non-MPI IPC runtime support, `reduce_on_stream` variants, a native `reduce_scatter` implementation, wave-level collective operations, and `team_split_2d` for refined team management.

Furthermore, two core libraries received targeted improvements: hipBLASLt includes a new GEMM kernel optimizer that allows teams to tune kernel selection locally, without exposing proprietary model weights. rocSPARSE adds enhancements for sparse workloads, including a smart SPMM selector, CSC triangular solves, and Blocked-ELL DenseToSparse conversion.

Developer Tools and Ecosystem

The profiling infrastructure has also seen substantial improvements. The Compute Profiler’s Roofline analysis now extends to GFX11xx (RDNA 3) architectures, providing performance-ceiling analysis—a critical tool for determining if a kernel is compute-bound or memory-bandwidth-bound—for the first time on Radeon hardware. ROCprofiler-SDK can now trace data in the Systems Profiler, allowing developers to observe communication activity alongside GPU compute activity for complete performance understanding. ROCm Optiq 1.0 is available as the general-purpose unified visualization and analysis environment, combining data from the ROCm Systems Profiler and the Compute Profiler into a single tool for seamless analysis of large-scale AI and HPC workloads.

From a developer experience standpoint, ROCm 10.0 packages are now delivered with Address Sanitizer (ASAN) instrumentation, making the process of catching memory-safety bugs significantly easier by treating ASAN builds like standard package installations. Additionally, the standard packaging mechanism for installed packages now utilizes `RPATH` instead of `RUNPATH`. This change ensures that ROCm binaries reference the libraries against which they were built and validated, minimizing conflicts when multiple versions of ROCm are present on a single machine.

For Windows users, the HIP SDK has been retired. Its functionality is absorbed by the ROCm Core SDK, meaning that both Windows and Linux now share the same SDK definition and release schedule, simplifying the development lifecycle. While ROCm on Windows currently ships as a portable tarball, native Windows installers are scheduled for later in 2026.

Summary

ROCm 10.0 represents a decade-long maturation of AMD’s open GPU software platform. Beyond the core improvements delivered by TheRock and the new packaging model, the arrival of ROCm.AI—with its integrated CLI, AMD Skills, and Hyperloom—marks a fundamental shift in how developers and automated agents interact with and build upon AMD hardware. The release also includes major investments in advanced communication capabilities, most notably the significant updates to RCCL and rocSHMEM, ensuring the platform is robust for massive, real-world deployment scales.

Written by

Hue

Hue is obsessed with GPU benchmarks and checking her crypto portfolio between gaming sessions. She writes about PC tech, games, and crypto.