Access all ROCm™ developer resources here — documentation, training, webinars, latest blogs, and more.

AMD ROCm

What is AMD ROCm software: Programming AMD GPUs

ROCm is an open, production-grade software platform with drivers, libraries, and developer tools for programming AMD GPUs.

It supports HPC and AI workloads, enabling developers to train and serve models at scale using familiar frameworks like PyTorch. 

ROCm Software for AI is Easier Than You Think
Learn More about ROCm Software
ROCm Optimized GPU Software Stack

What's New

AMD Ryzen AI Halo

Run AI Workloads Locally with AMD AI Playbooks

AMD AI Playbooks are free, step-by-step guides covering inference, fine-tuning, agent workflows, and distributed inference - all running locally on AMD client platforms. The AMD Ryzen™ AI Halo ships with pre-configured ROCm support, so you can go from any playbook straight to a running workload with no cloud dependency.

Explore AMD Playbooks

Start Building With ROCm

Find the Right GPU for Your Workload 

AMD Instinct GPUs are built for training AI models, running large-scale inference services, and HPC simulations.  

AMD Radeon GPUs are built for AI-powered gaming and real-time graphics, creative workflows, and local LLMs. 

AMD GPUs for AI: How to Choose
AMD GPUs for AI: How to Choose infographic

Move from CUDA to ROCm with Few to No Code Changes

You can run many workloads on AMD GPUs — without changing the code. When changes are needed, the HIPIFY tool helps you convert existing CUDA code to ROCm-compatible HIP code with minimal disruption. 

CUDA to ROCm: Which Migration Path is Right for You?
Move from CUDA to ROCm with Few to No Code Changes infographic

Join the AI Developer Program

Access free AMD Developer Cloud credits, exclusive training, monthly AMD GPU sweepstakes, and community support designed to support your AI development work.

Join Program

Frequently Asked Questions

ROCm and CUDA are both platforms for programming GPUs, but they differ in their approach.

ROCm is a full open-source, production-ready software stack compatible with a broad range of AMD GPUs. It supports robust tools, AI models, and frameworks, including PyTorch, TensorFlow, JAX, and more. The platform offers the flexibility to customize, modify, and optimize for specific environments and requirements — and is supported by a vibrant open ecosystem for faster innovation.

CUDA is NVIDIA’s proprietary platform, tightly integrated with its hardware. While CUDA is mature, its closed ecosystem can lock in developers and limit choice.

Many AMD Instinct GPUs, AMD Radeon GPUs, and iGPUs in select AMD Ryzen™ AI processors are compatible with ROCm software, giving developers a broad range to choose from. Check out the matrices below to see detailed GPU compatibility information for compute and graphics workloads. You’ll also find supported open-source frameworks, libraries, and tools and system requirements.

Compatibility Matrix: Instinct and Radeon for ROCm Compute Workloads >

Compatibility Matrix: Radeon and Ryzen for ROCm Graphics Workloads >

ROCm is primarily used for GPU-accelerated AI and HPC workloads, especially deep learning training, fine-tuning, and inference across frameworks like PyTorch, TensorFlow, and JAX.
Key use cases include:

  • Distributed training across multiple GPUs
  • Inference serving via optimized engines like vLLM and SGLang
  • Model optimization using Hugging Face and other open-source ecosystems

Developers can run existing codebases on AMD GPUs with minimal changes.

The AMD Developer Community forum and the AMD Developer Discord server allow you to connect directly with other developers and AMD experts to discuss questions and help troubleshoot issues. You can also submit a bug on GitHub.