AMD ROCm

AMD ROCm (originally standing for Radeon Open Compute platform) is an open-source software stack developed by Advanced Micro Devices (AMD) for graphics processing unit (GPU) programming. ROCm spans general-purpose computing on GPUs (GPGPU), high-performance computing (HPC), and heterogeneous computing, and offers several programming models including HIP, OpenMP, and OpenCL. It is distributed under permissive open-source licences and provides the compilers, runtimes, and libraries needed to run AI and HPC workloads on AMD hardware.^[1][2]

History and Background

ROCm traces its origins to AMD's earlier GPU computing efforts, including the Close to Metal, Stream, and HSA (Heterogeneous System Architecture) initiatives. The first public release of the platform, ROCm 1.3, was announced at the SC16 supercomputing conference in November 2016, under the name Radeon Open Compute platform. The name was later shortened to "ROCm" because "Open Compute" was a registered trademark of another organisation.^[1]

AMD has maintained ROCm as an open-source project throughout its history, with the source hosted on GitHub. In July 2023 AMD announced a new HIP SDK designed to help democratise GPU computing, and the platform has followed a rapid release cadence — roughly eight to nine feature releases per year — with ROCm 7.14.0 released on 16 July 2026.^[1][3]

Key Concepts and Technology

HIP: The Heterogeneous Interface for Portability (HIP) is ROCm's C++ runtime API and kernel language, designed to provide an interface similar to NVIDIA's CUDA so that developers can write portable GPU code that runs on both AMD and NVIDIA hardware. HIP is central to ROCm: PyTorch and other major frameworks use HIP to support AMD GPUs, and the hipcc compiler, based on LLVM, builds HIP kernels for AMD hardware.^[1][2][4]

Software stack: ROCm bundles the AMD GPU kernel driver, an LLVM-based compiler toolchain, maths and communication libraries (including rocBLAS and rocFFT), profiling and debugging tools, and support for OpenMP and OpenCL, providing an end-to-end platform between AI frameworks and AMD silicon.^[2][3]

Framework compatibility: ROCm runs the major machine learning frameworks natively, including PyTorch, TensorFlow, JAX, ONNX Runtime, Triton, vLLM, and SGLang, with minimal code changes, and provides day-0 support for major open-weight model releases on AMD hardware. It also supports low-precision data types including FP8, FP6, and FP4 to improve inference throughput per watt on supported GPUs.^[3]

Licensing: ROCm is free and open-source software (with the exception of GPU firmware blobs) distributed under various licences, with the core stack under the MIT licence.^[1][2]

Applications and Impact

ROCm is used for GPU-accelerated high-performance computing, AI training and inference, scientific computing, and computer-aided design. It is the software foundation for AMD's Instinct data centre accelerators, enabling AMD GPUs to serve as an alternative to NVIDIA systems for AI workloads. The growing maturity of ROCm — including native support in PyTorch and the vLLM inference server — has made AMD hardware an increasingly credible option for AI deployment, particularly for organisations seeking open-source software and supply-chain diversity.^[2][3][4]

>See Also

• AMD
• CUDA
• PyTorch
• GPU Cluster
• vLLM
• NVIDIA

References