AMD launches ROCm 10 to ease AI software migration from CUDA

AMD has released ROCm 10, a new version of its open software platform for GPU computing and artificial intelligence development. Alongside this release, the company has made ROCm.AI generally available, providing a suite of tools designed for the development, execution, and optimization of AI applications.

A primary focus of the update is simplifying the migration process for users currently utilizing NVIDIA’s CUDA platform. AMD utilizes HIP, a programming layer with an interface similar to CUDA, and HIPIFY, a tool capable of automatically translating portions of existing CUDA code into a format compatible with ROCm. For applications written in PyTorch, the transition is further streamlined as the ROCm version of PyTorch maintains the torch.cuda interface, often requiring minimal code changes.

The ROCm 10 ecosystem includes containers for tools such as vLLM and SGLang, Python packages, and a modular Core SDK. Additional features include Primus, a system for preparing and monitoring AI model training, and Hyperloom, an optimization tool that analyzes application execution to identify and address bottlenecks. AMD also introduced improvements to the RCCL communication library to enhance performance across multiple accelerators and server nodes.