CUDA

CUDA (Compute Unified Device Architecture) is NVIDIA's parallel-computing platform and the dominant API for GPU-accelerated AI. Every major AI framework — PyTorch, TensorFlow, JAX, llama.cpp, vLLM, ExLlamaV2 — has CUDA as its primary or best-supported backend.

CUDA's incumbency creates the "NVIDIA tax": even when AMD GPUs have comparable hardware (more VRAM at lower price), the software ecosystem leans NVIDIA. ROCm (AMD's CUDA equivalent) has improved significantly on Linux but still trails on Windows and on training workloads.

Practical implication for buyers: if you want to run AI without troubleshooting, get an NVIDIA card. If you're comfortable on Linux and willing to file occasional GitHub issues, AMD's price-per-VRAM-GB advantage can be real. Apple Silicon sidesteps the question entirely with the Metal/MLX path.

A buyer comparing a used RTX 3090 (24GB, roughly $700) against an RX 7900 XTX (24GB, roughly $800) for local LLM inference picks the 3090 despite the worse price-per-VRAM-dollar ratio, because every quantization tool they plan to use — ExLlamaV2, AWQ, bitsandbytes — assumes CUDA as the primary backend, and ROCm support for those tools on Windows is inconsistent enough to cost real debugging time. A Linux-only shop willing to build llama.cpp with ROCm and stick to GGUF (which has solid ROCm support) could make the AMD card work fine, but for someone who wants pip install to just work, the NVIDIA tax is the price of avoiding a week of troubleshooting. Anyone checking hardware fit before buying can run the target model through /will-it-run to see which backend paths are actually validated.