ROCm
AMD's open source GPU compute stack. Competitor to Nvidia's CUDA. Launched 2016. Powers AMD Instinct MI300X accelerators used for Llama training at Meta and among LLM hosters.
HIP is AMD's CUDA compatibility API — code compiles to either AMD or Nvidia. PyTorch and JAX have officially supported ROCm since ~2022. Real-world adoption has been rough: driver stability, narrow hardware support (only high-end cards), less developer tooling. The MI300X (192 GB HBM3) became, in 2024, the first AMD card that actually competed in LLM inference vs the H100. Lower price per byte of VRAM. Consumer GPUs (RX 7900 XTX) have support but marginal.