ROCm
AMD's open-source answer to CUDA — enables AI inference on Radeon GPUs.
Definition
ROCm (Radeon Open Compute) is AMD's open-source software stack for GPU computing. It provides CUDA compatibility layers (HIP), math libraries, and ML framework support (PyTorch-ROCm). ROCm enables AI inference and training on AMD RDNA 3/4 and CDNA GPUs.
Why It Matters
Medium. ROCm compatibility has improved significantly in 2024-2025. Tools like llama.cpp, Ollama, and PyTorch now officially support ROCm — making AMD RX 7900 XTX (24GB VRAM) a viable local AI option at lower cost than the RTX 4090. However, debugging is harder than CUDA.
Real-World Example
To run Ollama on an AMD RX 7900 XTX with ROCm: install ROCm 6.x drivers, then 'ollama run llama3.3' works identically to the NVIDIA experience — just with AMD's compute backend.
History of ROCm
AMD open-sourced ROCm in 2016 as its response to CUDA's ecosystem lock-in. Progress was slow until 2022-2023, when ML framework maintainers began officially supporting ROCm wheels. AMD's MI300X (2024) with 192GB HBM3 made ROCm critical for datacenter-scale inference.