CUDA at 20: How GPU Programming Became Part of Scientific Computing and AI
- Cargo
- Machine Learning
- Shipped
- Shipper
- Brett Murray
- Readers
- 647
From molecular simulations to AlexNet and AlphaFold, CUDA helped make GPU acceleration a practical tool across science and AI.

CUDA’s significance comes from a shift in who could put GPUs to work. Introduced as a programming platform for general computing, it gave researchers and developers a way to use graphics processors for workloads such as simulation and machine learning. Two decades later, NVIDIA says more than 11 million developers build on CUDA, and its influence reaches from scientific applications running on supercomputers to software for protein structure prediction.
The platform’s history is also a story about tools and research reinforcing one another. As developers built programs and libraries for GPU computing, researchers could take on workloads that benefit from parallel processing. Their results, in turn, helped show other fields where accelerated computing could fit. The examples span molecular biology, image recognition and modern AI.
One programming model, more kinds of work
NVIDIA announced CUDA in November 2006 and introduced it to the scientific community at SC06. Its high-level programming model let scientists and developers use GPUs for tasks beyond graphics. That mattered for applications such as scientific simulation, where many calculations can be carried out in parallel and faster computation can make more detailed or frequent analysis practical.
The broader change was not simply putting a faster processor into a computer. Developers needed a way to express work for a GPU, along with software that helped them apply it to specific tasks. CUDA offered that programming foundation. Over time, libraries and other tools built around the platform gave teams more ways to use GPU acceleration without starting every project from scratch.
That software layer now includes more than 1,000 CUDA-X libraries, according to NVIDIA. The number points to an ecosystem with components aimed at different stages and kinds of computing work. For teams, these libraries can provide ready-made building blocks; the practical value depends on whether a library supports the workload and hardware they use.
Molecular biology meets supercomputing
At the University of Illinois Urbana-Champaign, biophysicist Klaus Schulten’s team helped bring GPU computing into molecular biology through Nanoscale Molecular Dynamics and Visual Molecular Dynamics. The tools support molecular simulation and visualization, connecting computation with the study of biological structures.
A notable example came in 2013, when researchers from Illinois, the University of Pittsburgh and Vanderbilt University worked to resolve the HIV capsid’s atomic structure. They combined experiments with simulations run on Blue Waters, then among the world’s most powerful supercomputers. This case illustrates how accelerated computing can contribute to a research workflow alongside experimental evidence: the computer models help explore molecular behavior, while experiments provide another line of evidence.
The result also highlights a practical point about scientific computing. A platform’s impact depends on more than its programming model. Researchers need suitable software, access to computing resources and methods that connect computational output to the scientific question. CUDA-supported tools formed part of that larger effort.
GPU training helps expose deep learning’s potential
CUDA’s role in AI became particularly visible with AlexNet. University of Toronto researchers Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton used CUDA to train the image-recognition system on two NVIDIA GPUs. AlexNet won the 2012 ImageNet competition, helping demonstrate the potential of GPU computing for deep learning.
The example links a concrete engineering choice to a wider research shift: GPUs made it possible to train a prominent image-recognition model using parallel hardware, and the result helped bring attention to deep learning’s capabilities. It doesn’t mean that hardware alone produced the advance. The model, training approach, data and software all mattered. But CUDA gave the team a way to use GPUs for the demanding computation involved.
That pattern continues in scientific AI. Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for work on protein structure prediction. Their team’s publicly released AlphaFold 2 software uses CUDA for GPU execution. Here, the platform supports a system designed to predict protein structures, illustrating how GPU computing can be part of software used to tackle scientific problems.
The ecosystem is the long-term story
These examples show different forms of impact: simulation and visualization for molecular research, GPU training for image recognition, and accelerated execution for protein prediction. They don’t establish that every scientific or AI workload benefits equally. The fit depends on the task, the available hardware and the software that connects an application to the GPU.
Still, the range helps explain why CUDA’s 20-year milestone extends beyond a single programming interface. Developers have built tools that let researchers and engineering teams apply GPU computing across fields. NVIDIA points to uses that include medicine research, factory design, autonomous vehicles, financial market modeling and quantum computing. Those areas have distinct requirements, so the platform’s presence alone doesn’t determine what a system can achieve. It does offer a shared base for developers to build on.
For organizations assessing GPU acceleration, the useful questions are specific: whether their workload can benefit from parallel computation, whether the required CUDA libraries support it, and what computing resources deployment will demand. CUDA’s record shows how a programming platform can help move GPU use from a specialized capability into a broad set of research and development workflows. Its next chapter will depend on the software and discoveries that developers build with it.