Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

CUDA at 20: Nvidia marks two decades of GPU computing at SC26

Nvidia is marking the twentieth anniversary of CUDA this week at SC26, honoring the researchers, developers and engineers who built on its GPU programming model, according to a company blog post published November 16. The retrospective frames two decades of CUDA as the story of how a graphics platform became the default substrate for scientific computing and AI.

Published
CUDA二十周年:英伟达在SC26回望GPU计算如何成为科学与AI的地基
Image source: blogs.nvidia.com

Nvidia is marking two decades of CUDA this week at SC26, using the anniversary to honor the researchers, developers and engineers who turned GPU computing into a foundation for modern science and artificial intelligence, according to a company blog post published November 16.

The occasion is a retrospective rather than a product launch. CUDA is Nvidia's parallel computing platform and programming model, the layer that lets developers treat a GPU as a general-purpose processor instead of a device that only draws pixels. Over twenty years that layer became the default on-ramp to compute for both simulation and deep learning.

The blog tells the story through people, from the authors of the earliest CUDA programs to today's AI developers. That framing is itself revealing: the platform's advantage has never been raw throughput alone but the surrounding stack of compilers, libraries, profilers and, above all, the working habits developers build over years.

SC26 gives the anniversary a natural stage. SC is the high-performance computing community's annual gathering, and CUDA's lifespan overlaps almost exactly with the shift of GPUs from add-in accelerators to the center of the machine. Today's debates about power, bandwidth and interconnect in AI data centers still resolve into questions about the abstractions this programming model defined.

That is why an anniversary post belongs in a news feed at all. Silicon turns over every couple of years; programming models last far longer. Code, tuning knowledge and team skills accumulated on one platform raise the cost of moving to another, and that accumulation is precisely what Nvidia is drawing attention to with the number twenty.

What to watch next is whether Nvidia uses SC26 to extend that ecosystem onto new hardware and interconnect designs, and whether alternatives — in-house compiler stacks or open standards — can close the gap on real scientific workloads. The anniversary is a marker rather than an endpoint: the next decade of compute competition is likely to be decided less by a single chip than by who defines how developers write their code.

Why it matters

Two decades of CUDA show that AI compute advantage rests on the time developers have already invested in a platform, not on chip specifications alone. For challengers, the real barrier is switching cost rather than a point-to-point performance gap.

NVIDIACUDAGPU
Back to realtime news

Nearby Updates

All

11/17, 00:00

Nvidia expands CUDA-X with cuPhoton, cuPotential and cuSpa for AI research agents

Nvidia expanded its CUDA-X libraries with cuPhoton, cuPotential and cuSpa, aimed at research in astrophysics, materials science and single-cell biology. The company says the speedups let scientists analyse larger datasets and model more complex systems, while giving AI agents callable tools for discovery.

11/17, 00:00

Nvidia unveils Rubin HPC, a 50,000-core-plus GPU built for science agents

At SC26, Nvidia announced Rubin HPC, a new GPU in its Vera Rubin platform aimed at simulation and data processing for science, engineering and analytics, with more than 50,000 CUDA cores on a single chip. Nvidia says the card scales data movement alongside compute, paired with sparse-data acceleration and seventh-generation NVLink, to shorten the loop from an agent's hypothesis to a simulated result.

11/17, 00:00

Nvidia's Vera Rubin debuts in MLPerf with a training sweep and 10x agentic efficiency gain

Nvidia published Vera Rubin's first MLPerf results, claiming leadership across training, post-training and agentic inference, and a sweep of every MLPerf Training v6.1 benchmark. In its first MLPerf Endpoints submission, the platform reports more than 10x generational gains in throughput per watt for agentic AI.

11/17, 00:00

NVIDIA Extends Slurm to Put Quantum Processors Inside Supercomputing Workflows

NVIDIA is expanding Slurm's support for quantum processors so supercomputing centres can schedule quantum hardware alongside their existing scientific workloads. The move lands as nearly 30 centres prepare to deploy more than 50 quantum processors, a roughly sevenfold rise in deployments over five years.