AMD launched ROCm 10 to mark the open‑software stack’s tenth anniversary, alongside general availability of ROCm.AI for developers. Unveiled at Advancing AI 2026, ROCm.AI is an AI‑native development platform designed to accelerate building, deploying and optimizing workloads on AMD hardware.

ROCm.AI unifies three components: ROCm Hyperloom, AMD Skills and the ROCm CLI, delivering AMD‑specific expertise and agent‑driven optimization into existing developer toolchains. AMD reports that on equivalent hardware, ROCm.AI‑equipped systems deliver average 3.3× higher inference performance and 2.4× better training throughput versus ROCm 7 via AI‑optimized kernels, memory handling and scheduling.

Autonomous Inference Optimization ROCm Hyperloom acts as an agentic engine for end‑to‑end inference optimization covering host code and GPU kernels. It profiles workloads, identifies bottlenecks, explores optimization paths, applies targeted code edits, then benchmarks and validates correctness.

In ROCm 10, Hyperloom extends support for AMD Instinct GPUs and adds compatibility with vLLM and SGLang inference frameworks. Developers can trigger optimizations across HIP, Triton and FlyDSL codebases; Hyperloom generates reports listing proposed changes together with performance projections. It runs standalone or integrates with AMD Skills.

Curated Expertise for AI Coding Agents AMD Skills injects validated AMD technical knowledge and workflows into AI coding agents including Claude Code, Cursor and Codex, supplying hardware‑specific guidance without context‑switching.

ROCm 10 expands the Skills catalog into three groups: local‑client AI workflows, cross‑stack diagnostics‑oriented workflows, and server‑side workflows for Instinct GPUs and EPYC CPUs covering serving, profiling and performance analysis. Previewed Advancing‑AI capabilities are now available via agent marketplaces and an open‑source GitHub catalog, with each skill undergoing structural and functional testing pre‑release.

Unified Command‑Line Management The ROCm CLI, currently in technology preview for ROCm.AI, offers a consistent command‑line surface to provision, manage and execute AI workloads on AMD hardware. The same commands work interactively, via AI agents or inside CI pipelines for system inspection, ROCm lifecycle management, model serving, diagnostics and runtime control.

Distributed as pre‑built Windows and Linux binaries with no pre‑existing ROCm dependency, it supports multiple side‑by‑side managed runtime environments including activation and rollback, plus integrated model‑serving orchestration. Current adapters cover Lemonade on select client hardware and vLLM for Instinct deployments.

Packaged with the CLI is ROCm Console (internally “dash”), providing real‑time telemetry for runtime health, model serving status, GPU utilization, HBM consumption, power draw and tokens‑per‑watt metrics on supported Instinct hardware. The preview delivers version‑agnostic operation starting from ROCm 7.13; native ROCm 10 support is planned.

Broader ROCm 10 Stack Updates These developer tools sit atop the revamped ROCm 10 stack featuring a more modular Core SDK, alongside updated libraries, compilers, framework integrations, model support, tuning enhancements and expanded hardware enablement.

Test Configuration Reference

Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!