AMD는 ROCm 10을 출시하고 ROCm.AI의 일반 공개(GA)를 발표하며, 오픈소스 AI 소프트웨어 스택의 10년 간의 성숙을 완성했습니다.
On August 31, 2026, AMD released ROCm 10, commemorating a decade of its software stack and bringing ROCm.AI to general availability. Initially unveiled at Advancing AI 2026, ROCm.AI is an AI-native software experience engineered to accelerate development velocity and optimize performance across AMD hardware.
ROCm.AI unifies three core developer experiences: AMD Skills, the ROCm CLI, and AMD Hyperloom. By embedding AMD-specific expertise and agentic workflows directly into familiar developer tools, the platform enables innovation to keep pace with AI advancements. Leveraging AI-driven optimization for kernels, memory management, and scheduling, systems configured with ROCm.AI deliver an average 3.3x improvement in inference and a 2.4x improvement in training compared to ROCm 7 on identical hardware.
**A New Developer Experience for AI on AMD**
The platform introduces a streamlined workflow supporting engineers across three key stages of the AI development lifecycle: optimizing inference performance, building with AMD expertise, and running AI workloads.
**Optimize End-to-End Inference with AMD ROCm Hyperloom**
As a core component of ROCm.AI, Hyperloom is an autonomous agentic system designed to optimize end-to-end inference workloads across both host code and GPU kernels. It automatically profiles workloads, identifies bottlenecks, explores optimization pathways, implements targeted adjustments, benchmarks outcomes, and validates both performance and correctness. ROCm 10 expands Hyperloom’s compatibility to AMD Instinct GPUs, adding support for vLLM and SGLang. Developers can direct optimizations across HIP, Triton, and FlyDSL, receiving detailed reports that outline proposed code modifications alongside measured or projected performance gains. Hyperloom integrates seamlessly via standalone workflows and AMD Skills, offering developers flexible methods to embed agentic optimization into their existing pipelines.
**Build with AMD Expertise**
AMD Skills injects curated AMD knowledge and validated workflows directly into leading AI coding agents, including Claude Code, Cursor, and Codex. This provides developers with AMD-specific guidance within their preferred environments. ROCm 10 significantly expands the AMD Skills catalog across three domains: client-native workflows for local AI and application integration; cross-stack workflows covering diagnostics, routing, replay analysis, and optimization; and server-native workflows tailored for AMD Instinct GPUs and AMD EPYC processors, encompassing model serving, profiling, and performance analysis. Skills initially previewed at Advancing AI are now accessible via the Claude Code, Codex, and Cursor marketplaces, as well as through an open catalog on GitHub. Every released skill undergoes rigorous structural and behavioral testing to ensure consistent, reliable workflows.
**Run with Simpler Workflows**
The ROCm CLI, currently offered as a Technology Preview within ROCm.AI, delivers a stable, unified command-line interface for provisioning, managing, and operating AI workloads on AMD hardware. Engineers can execute these workflows manually, leverage AI coding agents, or deploy them within continuous integration (CI) environments. A single interface consolidates system inspection, ROCm environment installation and management, model serving, diagnostics, component updates, and runtime control. Available as a prebuilt binary for both Windows and Linux, the CLI operates independently without requiring a prior ROCm installation. It provisions managed ROCm environments, supports multiple side-by-side runtimes with activation and rollback capabilities, and includes integrated model serving and engine management. Current adapters enable Lemonade on select AMD client systems and vLLM for AMD Instinct GPU serving.
Bundled with the CLI is the ROCm Console (previously known as “dash”), which offers real-time visibility into system status and workload activity. Users can track ROCm runtime health, model serving operations, GPU utilization, and benchmark telemetry, including critical metrics like high-bandwidth memory (HBM) usage, power draw, and tokens per watt on compatible AMD Instinct systems. As a Technology Preview, the ROCm CLI maintains a version-agnostic experience spanning ROCm releases starting from 7.13, with official ROCm 10 support slated for imminent release.
**Building on the ROCm 10 Software Stack**
These developer experiences are built upon the expanded ROCm 10 software stack, anchored by the ROCm Core SDK. The SDK establishes a more modular architecture for developing and deploying AI workloads across AMD platforms. Beyond the SDK, ROCm 10 introduces comprehensive updates to libraries, compilers, frameworks, developer tools, model support, performance optimizations, and hardware platform compatibility. For a comprehensive breakdown of the ROCm Core SDK—including libraries, compilers, tools, framework and model support, performance gains, and platform updates—readers are directed to the full ROCm 10 technical deep dive.
**About AMD**
AMD (NASDAQ: AMD) advances high-performance and AI computing to address the world’s most critical challenges. Today, AMD technology underpins billions of experiences spanning cloud and AI infrastructure, embedded systems, AI PCs, and gaming. Backed by a comprehensive portfolio of AI-optimized CPUs, GPUs, networking solutions, and software, AMD delivers full-stack AI solutions engineered for the performance and scalability required in a new era of intelligent computing. More information is available at www.amd.com.