CUDA (Compute Unified Device Architecture) is a parallel computing platform and programming model created by Nvidia, first released in 2006, that lets developers write general-purpose code to run on Nvidia GPUs rather than only graphics-rendering code. It gave programmers a relatively accessible way to harness a GPU's thousands of parallel cores for arbitrary computation, extending languages such as C and C++ with GPU-specific constructs.
History and purpose
Before CUDA, using a GPU for non-graphics computation required expressing a problem in terms of graphics operations, such as textures and shaders, a workaround few outside computer graphics found practical. Nvidia's release of CUDA in 2006 removed that barrier, and the platform matured over the following decade through adoption in scientific computing, computational finance, and, increasingly from the early 2010s, deep learning research. The 2012 success of AlexNet, trained using CUDA-accelerated GPUs, demonstrated deep learning's potential and drove a surge of research investment into CUDA-based GPU training, cementing the platform's role in the field.
Technical role and ecosystem
CUDA provides the low-level libraries, such as cuDNN for neural network primitives and cuBLAS for linear algebra, that most deep learning frameworks build on. Both PyTorch and TensorFlow, the two dominant deep learning frameworks, are written to run efficiently on CUDA, and the large majority of published deep learning research and production AI systems as of the mid-2020s run on CUDA-enabled Nvidia hardware. This created a powerful software moat: switching a large AI codebase from Nvidia GPUs to a competitor's hardware, such as AMD's ROCm platform or Google's TPU, typically requires significant engineering work, because so much tooling, from individual research code to production inference servers, is written and optimized specifically against CUDA.
Significance and criticism
CUDA is widely credited, alongside the GPU hardware itself, with being a decisive factor in Nvidia's dominance of the AI hardware market during the AI boom of the 2020s, since it is difficult for competitors to dislodge Nvidia's position by matching hardware performance alone, given how much of the software ecosystem, from research libraries to large language model training frameworks, assumes CUDA is available. Critics of this dynamic, including competing chipmakers and some AI labs, have described CUDA as a proprietary lock-in mechanism that raises costs and limits hardware choice across the industry, and have backed open alternatives such as OpenAI's Triton language, aimed at making it easier to write GPU code portable across vendors. Despite these efforts, as of 2025 CUDA remained the default assumption underlying most deep learning software, from academic research code to the infrastructure running frontier foundation models.