Trusence Every claim has a source
Last updated 17 November 2026 Search Türkçe
Infrastructure

NVIDIA marks 20 years of CUDA as its scientific computing platform reaches 11 million developers

The anniversary comes alongside new open tools for science and a quantum decoding method that cuts one computation from 12 weeks to 45 minutes.

NVIDIA is using its SC26 conference week to mark 20 years since it introduced CUDA, now relied on by more than 11 million developers, and to recognise researchers whose work built on it, including Illinois, Toronto and the AlphaFold team. Separately, Ian Buck used an opening address to unveil Nemotron for Science, an open model and dataset that researchers can run on local hardware, and cited an Argonne agent that proposed two new binding sites from a single prompt. He also outlined the Rubin HPC GPU, slated for 2028, with more than 50,000 CUDA cores, and a Vera CPU deal with DKRZ. On quantum computing, NVIDIA, IBM and Fermilab announced a co-design platform, and a new CUDA-Q decoding algorithm was said to cut one error-correction computation from 12 weeks to 45 minutes.

Why it matters

CUDA has moved from a 2006 release into a platform that more than 11 million developers now use across the computing stack, so its direction shapes how a large share of scientific software gets built and run. The new open model and dataset for chemistry put some AI research tools on local hardware rather than behind a remote service. The quantum result shows the platform reaching into error-correction research, where the time saved on one computation is the clearest sign of practical change so far.

Sources · 2