AMD ROCm
Open-source GPU computing stack of compilers, runtimes and libraries for running AI and HPC workloads on AMD hardware.
What it does
- Model training
- Model inference
- GPU programming
Features 5
- ROCm CLIUnified command-line interface to set up, manage and operate AI workloads on AMD hardware.
- ROCm Core SDKModular foundation for building, deploying and optimizing AI workloads across AMD platforms.
- ROCm HyperloomUses agentic AI to profile workloads, find bottlenecks and implement and validate optimizations.
- AMD SkillsCurated AMD knowledge and validated workflows delivered into AI coding agents for AMD-optimized development.
- ROCm.AIAI-native developer experience that accelerates ROCm development and optimization on AMD hardware.
How it compares
-
CUDA Toolkit
Rival GPU programming stacks for the same accelerator workloads; AMD's own documentation ships a CUDA-to-HIP porting path — the challenger documenting the route off the incumbent toolchain.
Porting CUDA code to HIP
Official documentation · 25 Aug 2026 -
Mojo
The startup touts its unified stack as a way to deliver up to 70% latency reduction and 80% cost savings compared to vendor-specific runtimes like CUDA and ROCm.
News reporting · 19 Sep 2026
Sources
- Pricing
-
Official documentation · 23 Aug 2026
AMD's own documentation states ROCm "is AMD's open-source GPU computing platform ... It is open source, cross-platform"; no paid tier of the stack itself is offered.
- What it does
-
Official documentation · 26 Aug 2026
Audit finding D-5: the ROCm⇄CUDA edge scored zero capability overlap because ROCm was tagged only model-training/inference while CUDA is tagged gpu-programming. ROCm is a GPU programming stack by AMD's own description — compilers, runtimes and a CUDA porting guide — so the term the vocabulary already had now sits on the record.
- Sold within
-
AMD's own page states "AMD ROCm™ is an open software platform that helps data center teams quickly set up, deploy, and scale AI and HPC workloads"; installation is via the free ROCm Developer Hub ('Get Started') with no price stated anywhere on the page.
- Interfaces
-
Official documentation · 14 Oct 2026
Page: 'The ROCm CLI provides a unified interface to set up, manage and operate AI workloads on AMD hardware' / 'The ROCm Core SDK provides a modular foundation for building, deploying and optimizing AI workloads across AMD platforms'.
- Intended for
-
Page headline: 'AMD ROCm software helps developers, researchers, and infrastructure teams turn AMD GPU capability into sustained application performance'
Also from AMD
-
AMD Developer Cloud Infrastructure service
An on-demand cloud service giving developers browser and API access to AMD Instinct MI300X GPU instances for AI, ML and HPC workloads.
-
AMD Helios Rackscale Solution Hardware
A rack-scale AI system combining 72 Instinct MI455X GPUs with EPYC CPUs and Pensando networking for frontier-model training and large-scale inference.
-
AMD Instinct Hardware
Data-centre GPU accelerators for training and serving models, sold in OEM servers and rack systems.
-
AMD Ryzen AI Software Platform
A development stack of compiler, runtime and tools for porting pretrained models onto the NPU and integrated GPU of Ryzen AI processors.
-
AMD Vitis AI Platform
An open-source AI inference development stack of compiler, optimised NPU IP and runtime for deploying deep-learning models on AMD adaptive SoCs, FPGAs and Alveo cards.
-
AMD Enterprise AI Reference Stack Platform
An open-source reference stack for running enterprise AI workloads at scale on AMD compute: open-source AI frameworks and generative models connected to an enterprise-ready Kubernetes platform, with AMD Inference Microservices — prebuilt inference containers bundling model, engine and optimised configuration for AMD hardware.
-
AMD Kria SOMs Hardware
System-on-module family for deploying edge and physical AI, with production SOMs and partner-validated carrier systems.
-
AMD Pensando AI NICs Hardware
AI NIC family (Pollara 400 at 400 Gbps, Vulcano 800 at 800 Gbps Ethernet) for scale-out inter-GPU communication in AI training and inference clusters, built on the Ultra Ethernet Consortium specification.
-
AMD Pensando DPUs Hardware
Data processing unit family (Salina, Giglio) offloading networking, security and storage services from host CPUs in AI data centres, programmable via the P4-based Pensando software stack.
-
AMD Versal AI Core Hardware
Adaptive SoC/FPGA family for AI inference and adaptive compute workloads.
-
AMD Versal AI Edge Series Gen 2 Hardware
Second-generation Versal AI Edge adaptive SoC family for embedded AI preprocessing, inference and postprocessing.
-
AMD Radeon AI PRO R9700 Hardware
Workstation graphics card with 32GB of video memory and 128 AI accelerators for running and developing AI models locally on a desktop.
-
AMD Ryzen AI Halo Hardware
Compact developer computer with Ryzen AI Max processors, up to 192GB of unified memory and pre-installed AI software for running and building with large models locally.
-
AMD Ryzen AI PRO Processors Hardware
Processors for commercial laptops and desktops that add a dedicated neural processing unit to the CPU and GPU so AI features run on the PC itself.
-
Four- to twelve-core embedded processors with deterministic latency for industrial, automotive and physical-AI workloads that accelerate edge AI on the device.
-
Single-chip CPU, GPU and NPU processors for autonomous systems and robots that run perception, reasoning and control for physical AI on the device.
-
AMD Versal AI Edge Series Hardware
First-generation Versal adaptive SoCs for automated driving, factory and healthcare systems, combining programmable logic, AI Engines for inference and a real-time processor.
-
GAIA Application
AMD's open-source desktop and terminal app of local AI agents for Ryzen AI PCs that answers questions from your documents, edits files, searches code and works with spreadsheets.
Something wrong here? Send a correction — quote this product id: amd-rocm.