NVIDIA marks CUDA’s 20th anniversary at SC26, honoring its scientific roots
The anniversary retrospective traces the software behind AlexNet and AlphaFold 2. NVIDIA says more than 11 million developers now build on the platform.
Loading page…
The anniversary retrospective traces the software behind AlexNet and AlphaFold 2. NVIDIA says more than 11 million developers now build on the platform.
Listen to this story
NVIDIA’s SC26 observance is a retrospective on CUDA’s role in moving GPU computing from scientific applications into deep learning, not an announcement of a new platform release. The company says more than 11 million developers now build on CUDA and counts over 1,000 CUDA-X libraries, underscoring the ecosystem built around the programming model. For AI builders, the examples show CUDA serving as infrastructure across different research workloads, while the anniversary itself offers no new product or adoption milestone.
NVIDIA announced CUDA in November 2006 and introduced it to the scientific community at SC06.
In 2012, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton trained AlexNet on two NVIDIA GPUs; it won that year’s ImageNet competition.
Researchers combined experiments with CUDA-accelerated simulations on the Blue Waters supercomputer to help resolve the HIV capsid’s atomic structure.
NVIDIA is marking CUDA’s 20th anniversary at SC26 this week, honoring the researchers who helped take graphics processors into scientific computing and AI. In a November 16 retrospective, the company traces that history through molecular simulations, AlexNet and AlphaFold 2, with a celebration wall at its conference booth.
The observance centers on the people who built applications and tools around CUDA, rather than a new version of the platform. NVIDIA’s account follows researchers working on problems as different as the structure of HIV and image recognition, showing how the same programming platform supported both scientific simulation and deep learning.
NVIDIA announced CUDA in November 2006 and introduced it to the scientific community at SC06. It gave scientists and developers a high-level programming model for using graphics processing units, or GPUs. That extended their work beyond graphics into demanding calculations for science, simulation and other applications.
The retrospective puts the developer community at the center of that expansion: people writing early CUDA programs, building tools and experimenting with new uses for accelerated computing. Its historical examples span supercomputer applications and visual effects as well as the AI systems that later became prominent users of GPU computing.
One milestone came in 2012, when University of Toronto researchers Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton used CUDA to train AlexNet on two NVIDIA GPUs. The image-recognition system won that year’s ImageNet competition. NVIDIA highlights the result as a demonstration of GPU computing’s potential for deep learning, linking the platform’s scientific origins to AI research.
Another example concerns the HIV capsid, the virus’s protein shell. In 2013, researchers at the University of Illinois Urbana-Champaign, the University of Pittsburgh and Vanderbilt University combined experiments with CUDA-accelerated simulations to help resolve its atomic structure. They ran the simulations on Blue Waters, then one of the world’s most powerful supercomputers.
That work followed efforts by Illinois biophysicist Klaus Schulten’s team to connect GPU computing with molecular biology. NVIDIA singles out Nanoscale Molecular Dynamics and Visual Molecular Dynamics, tools for simulating and visualizing molecules. The example concerns a combination of experimental evidence and computer simulation, rather than an AI model producing the scientific result on its own.
The retrospective also points to AlphaFold, which predicts protein structures. Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for protein structure prediction. Their team’s publicly released AlphaFold 2 software uses CUDA to execute work on GPUs, according to NVIDIA—a specific software role within the research system.
NVIDIA now says more than 11 million developers build on CUDA across the computing stack. It also counts more than 1,000 CUDA-X libraries. Those are company-supplied descriptions of the ecosystem’s current scale, presented alongside the anniversary history, not adoption milestones identified as having been reached at SC26.
For conference attendees, the anniversary has a physical destination: NVIDIA’s booth hosts the 20-Year CUDA Celebration Event Wall, honoring the community whose work the company is highlighting this week.
Loading discussion...
Join the conversation
Explain how you would share credit between research achievements and the tools enabling them.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.