AMD ROCm

ROCm is AMD's open-source software stack for running compute workloads, especially machine learning, on AMD GPUs. It includes low-level drivers, the HIP runtime and compiler tools for porting CUDA code, and a set of math and communication libraries (like MIOpen and RCCL) that mirror what CUDA provides on Nvidia hardware. The goal is to let developers train and serve AI models on AMD Instinct accelerators and select Radeon cards without being locked into Nvidia's ecosystem.

It's aimed at ML engineers, HPC teams, and infrastructure groups who want an alternative to Nvidia given pricing, supply constraints, or a preference for open tooling. PyTorch and JAX both ship official ROCm builds now, and major cloud providers and AMD's own Instinct MI-series deployments have pushed compatibility forward significantly compared to a few years ago, when ROCm support was spotty and often required patched builds or workarounds.

What sets it apart is that it's the most mature, actively maintained open path to CUDA-like functionality on non-Nvidia silicon. HIP's source-level compatibility with CUDA makes porting existing kernels more tractable than writing against, say, OpenCL or SYCL from scratch. It still lags CUDA in library breadth, third-party tooling, and general polish, and hardware support is narrower than Nvidia's lineup, but for teams running large training or inference jobs on AMD accelerators it's the standard stack rather than one of several options.