CUDA
Also called: CUDA cores
NVIDIA's computing platform for programming its GPUs for general-purpose tasks. CUDA cores are the compute units of these cards.
CUDA refers to two related things. First, it is a software platform launched by NVIDIA in 2007: a set of tools, libraries and drivers that lets developers use the GPU for general-purpose computing, not just for display. By extension, the basic compute units found by the thousands in NVIDIA graphics processors are called CUDA cores.
Most artificial intelligence software relies on CUDA: PyTorch, inference engines such as vLLM or Ollama, and many scientific computing tools. This is one of the reasons for NVIDIA's dominant position in AI. CUDA cores handle a large number of operations in parallel, while Tensor Cores, found on recent cards, specifically accelerate the matrix multiplications at the heart of neural networks. The installed drivers and CUDA version must be compatible with the software you use.
On a product page, the number of CUDA cores gives an idea of computing power, but it only compares well between cards of the same generation. For AI, also look at the amount of VRAM, the memory bandwidth and the Tensor Core generation. In practical terms, check that your operating system and drivers support the card.
Our NVIDIA H200 page presents a data center GPU designed for the most demanding CUDA workloads.