# CUDA CUDA is a programming model, a compiler, a runtime, and a very large collection of libraries that together let ordinary programmers run computation on accelerators. Its significance in this collection is strategic as much as technical: the hardware position that dominates AI computing rests on nearly two decades of software accumulation, and that software is the asset competitors find hardest to match. > [!note] The one-sentence version > A way to write accelerator code in a familiar language, plus eighteen years of libraries that nobody wants to rewrite. ## What it provides - **A programming model.** Code is written as a **kernel**, a function executed by thousands of threads at once, with threads grouped into blocks that share fast local memory. The programmer expresses parallelism directly, and the hardware schedules it. - **A compiler and runtime.** Source compiles to an intermediate representation and then to the instruction set of a specific accelerator generation, which is how a program built today runs on hardware released later. - **Memory management.** Explicit transfers between host and accelerator memory, pinned host memory for fast transfer, and unified addressing so that one pointer is valid on both sides. - **Streams and events.** Independent queues of work that overlap computation with data movement, which is how a well-written program keeps the device busy continuously. - **Libraries.** The part that actually matters in practice: dense and sparse linear algebra, fast Fourier transforms, random number generation, deep learning primitives including the convolution and attention kernels, and the collective communication library that implements group operations over the accelerator fabric; see [[wiki/Collective Communication|Collective Communication]]. ## Why the ecosystem is the position A competing accelerator may match arithmetic throughput and memory bandwidth, and it still has to be reached through software. The practical requirements are large: a compiler of comparable quality, kernels for every operation a framework calls, performance tuning for each generation, a collective library that handles the fabric, profiling and debugging tools engineers trust, and compatibility maintained as frameworks change underneath. Each of those took years, and the result is that frameworks, research code, tutorials, and the habits of a generation of engineers were written against one platform. The industry's considered judgement, stated by analysts and by competitors alike, is that processors and accelerators have converged in capability while the software ecosystem remains ahead; the secret is the software. That assessment is worth holding precisely because it indicates where competition will actually be decided. The practical counter-strategy has been to compete at the framework level rather than the kernel level: if researchers write in a framework and the framework supports several backends, the kernel library beneath becomes an implementation detail; see [[wiki/PyTorch|PyTorch]] and [[wiki/ROCm|ROCm]]. ## Relationships - **programs:** [[wiki/Accelerator|Accelerator]] - **supports:** [[wiki/PyTorch|PyTorch]] · [[wiki/vLLM|vLLM]] · [[wiki/Collective Communication|Collective Communication]] - **uses:** [[wiki/NVLink|NVLink]] · [[wiki/PCIe|PCIe]] - **competes with:** [[wiki/ROCm|ROCm]] - **determines:** [[wiki/Depreciation of Compute Assets|Depreciation of Compute Assets]] through support lifetime - **part of:** [[wiki/Software and Orchestration|Software and Orchestration]] - **source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] - **collection route:** [[collections/American Infrastructure|American Infrastructure]] · [[collections/American|American Collection]] ## Evidence posture - **Established:** the kernel, block, and thread programming model; intermediate representation compilation permitting forward compatibility across generations; explicit and unified memory management; streams and events for overlap; the library set spanning linear algebra, transforms, deep learning primitives, and collective communication. - **Established (datestamped):** as of October 2026, independent analysts assess the software ecosystem as the leading differentiator while processor and accelerator hardware capability has converged. - **Strongly indicated:** framework-level abstraction as the effective competitive route for alternative accelerators; multiple backends ship in major frameworks and share outcomes are still developing. ## Sources - [Helios Is AMD's First AI System To Rival Nvidia Vera Rubin](https://www.cnbc.com/) — CNBC, October 2026; analyst assessment that processors and accelerators are on par while the established software ecosystem remains ahead, and that the decisive element is software. - Public accelerator programming model and library documentation, as of October 2026.