ABAB Pedia

NVIDIA CUDA

CUDA is NVIDIA’s parallel computing platform and programming model for its graphics processors. Developers express parallel tasks through general-purpose programming languages. Compilers, runtime software and libraries allow computation-intensive work to run on the GPU alongside a host CPU.

Contents7 sections
Key facts

The GeForce 8800 and general-purpose computing

NVIDIA’s GeForce 8800 generation used a unified processing architecture. Earlier attempts to run general numerical work through graphics APIs often required programmers to represent data as textures and computations as graphics operations. CUDA exposed threads, memory and computation directly to programmers, giving the same hardware a more accessible role beyond graphics.⁠[1][2]

The first toolkit and the host–device model

NVIDIA’s archive dates CUDA Toolkit 1.0 to June 2007. The first guide describes the CPU side as the host and the GPU as the device. The host launches a kernel, which many threads execute on different data. Threads form blocks, and blocks form a grid. Threads within a block can coordinate and use shared memory. This organization lets programmers describe parallel work while leaving assignment to individual processing units to the system.⁠[2][3][4]

Developing and maintaining the toolkit

Archived documentation and downloads include release 6.0 in April 2014 and 10.0 in September 2018. The toolkit includes a compiler for device code, performance-analysis tools and libraries for operations such as linear algebra. Separate version documentation allows developers to match APIs and hardware requirements to an installed environment.⁠[3][5]

CUDA 13 and Blackwell support

CUDA Toolkit 13.0 expanded support for Blackwell GPUs, which had first been supported in 12.8. Alongside the existing thread-parallel model, the release laid groundwork for expressing computation in tiles. The changes concerned both supported devices and the abstractions available to programmers.⁠[6][3]

The 13.5 toolkit and distribution of samples

NVIDIA’s archive lists CUDA Toolkit 13.5.0 in November 2026 with version-specific documentation. NVIDIA’s license governs the toolkit, with separate terms for certain components. NVIDIA also maintains a separate GitHub repository of samples for learning the model and testing a development environment.⁠[3][7][8]

Related Organizations, Websites, and Public Accounts

CUDA Toolkit documentation, tools and downloads. https://developer.nvidia.com/cuda/toolkit (opens in a new window)⁠[5]

CUDA samples maintained by NVIDIA on GitHub. https://github.com/NVIDIA/cuda-samples (opens in a new window)⁠[7]

NVIDIA’s official CUDA platform site. https://developer.nvidia.com/cuda (opens in a new window)⁠[9]

Sources9 sources

  1. NVIDIA GeForce 8800 GPU Architecture Overview (opens in a new window)nvidia.com↑
  2. NVIDIA CUDA Programming Guide, Version 1.0 (opens in a new window)developer.download.nvidia.com↑
  3. CUDA Toolkit Archive (opens in a new window)developer.nvidia.com↑
  4. 2.1. Intro to CUDA C++ (opens in a new window)docs.nvidia.com↑
  5. CUDA Toolkit (opens in a new window)developer.nvidia.com↑
  6. What’s New and Important in CUDA Toolkit 13.0 (opens in a new window)developer.nvidia.com↑
  7. NVIDIA/cuda-samples: Samples for CUDA Developers (opens in a new window)github.com↑
  8. End User License Agreement: NVIDIA CUDA Toolkit (opens in a new window)docs.nvidia.com↑
  9. CUDA Platform for Accelerated Computing (opens in a new window)developer.nvidia.com↑