Glossary Updates6 new terms added to the glossaries · October 11, 2026, 13:38 CEST
AI TechDocKnowledge

Glossary · 5 · Hardware SDKs and devices

CUDA

Also known as: CUDA Toolkit, CUDA platform

In computing, CUDA is NVIDIA's parallel computing platform and programming model for running general-purpose computations on NVIDIA graphics processing units; its CUDA Toolkit, first released in 2007, provides a compiler, runtime, GPU-accelerated libraries and debugging and profiling tools.

  • Hardware SDK
  • Vendor product
  • AI

In one sentence

CUDA is NVIDIA's parallel computing platform and programming model for running general-purpose code on its GPUs, with compiler, libraries and tools.

Example

A research team writes a CUDA kernel in C++ that computes many matrix tiles in parallel and calls NVIDIA's math libraries for the rest, so the training job runs on the GPU instead of the CPU.

How it applies

  • Hardware SDKs: CUDA is a hardware SDK for accelerators. The CUDA Toolkit includes GPU-accelerated libraries, debugging and optimization tools, a C/C++ compiler (NVCC) and a runtime library. Code that runs on the GPU is written as kernels, functions that are executed by many GPU threads in parallel.
  • Libraries and languages: Most applications use CUDA indirectly through libraries for linear algebra, signal processing or neural networks and through frameworks for machine learning (ML). NVIDIA's programming guide covers CUDA C++ and CUDA Python; other languages reach CUDA through language bindings.
  • AI workloads: Training and running neural networks consist largely of parallel arithmetic, which is why GPU platforms such as CUDA are widely used for training and inference. The vendor itself is described under NVIDIA.
  • Developer and platform SDKs: Cloud SDKs provision GPU instances and container images with CUDA preinstalled; host applications expose GPU acceleration through plug-ins. The CUDA version, the driver version and the GPU generation have to fit together.
  • Documentation: State the CUDA Toolkit and driver versions an SDK or application was built and tested with, the supported GPU architectures and what happens on systems without a compatible GPU.

CUDA vs. vendor-neutral GPU interfaces

CUDA targets NVIDIA GPUs only. Vendor-neutral standards such as OpenCL and SYCL are specified by the Khronos Group as open, royalty-free, cross-platform interfaces for heterogeneous processors from several vendors. Code written for CUDA does not run on other vendors' GPUs without porting or translation layers.

Keep in mind

CUDA is a vendor product. Its features, supported GPU generations, languages and licensing terms change with each release. The CUDA Toolkit archive lists version 1.0 as June 2007; check the current NVIDIA documentation for everything else. Facts here are as of October 11, 2026.

External references

By knowledge.aitechdoc.world · Published October 11, 2026 · Last reviewed

Source: NVIDIA, “CUDA Programming Guide” and “CUDA Toolkit Archive”

Definitions follow the cited standards and specifications. Where a source is a copyrighted publication, such as an ISO, IEC or EN standard, the definition is a close paraphrase, not a verbatim quotation, so as not to infringe copyright. We recommend reading the original publication. The sections “How it applies” are editorial commentary by AI TechDoc Knowledge and are not part of any standard.

Seen a mistake? Send us a note!