ROCm
AMD's open-source GPU compute (GPGPU), HPC and AI software stack — drivers, the HIP programming model and libraries (rocBLAS, MIOpen, RCCL).
Description
ROCm is AMD's open-source software stack for general-purpose GPU computing (GPGPU), high-performance computing (HPC) and artificial intelligence. It is a direct alternative to NVIDIA CUDA on AMD hardware.
The stack comprises drivers and a kernel layer (amdgpu/ROCk), a runtime (ROCr), the HIP programming model and a set of optimized libraries — including rocBLAS (linear algebra), MIOpen (deep learning primitives), RCCL (multi-GPU communication), rocFFT and rocSPARSE.
ROCm supports popular machine learning frameworks such as PyTorch, TensorFlow, ONNX and JAX, and targets AMD Instinct accelerators (MI series) and Radeon GPUs. The HIPIFY tool enables automatic conversion of CUDA code to HIP.
The software is developed primarily on Linux (with Windows support for selected components) and released under open licenses (including MIT). Its first release was on 14 November 2016.