Tech • AI • Robotics • Game

ENFR

NVIDIA SC26 Special Address

9.5/10
NVIDIANVIDIANovember 18, 2026 at 06:36 PM48:11
Audio player

⏱ ≈ 3 min read · 48-min video — 45 min saved

TL;DR

Nvidia used a major supercomputing gathering to position AI, simulation, quantum computing and new desktop-scale systems as a unified scientific stack, while unveiling tools and hardware aimed at faster research from cancer biology to black hole imaging.

KEY POINTS

CUDA’s expansion into a broad scientific platform

Nvidia traced its scientific computing push back to 2006, when CUDA first brought the C programming model to GPUs. The platform has since grown into more than 1,000 SDKs and CUDA-X libraries spanning math, genomics, climate, robotics, quantum computing and drug discovery. Nvidia said those libraries have surpassed 1.2 billion downloads, underscoring how GPU computing has moved from niche acceleration to a core research platform.

South Pole neutrino science processed on site

The IceCube observatory beneath the South Pole uses more than 5,000 sensors buried up to 2.5 kilometers deep to detect rare neutrino interactions in ice. The instrument generates roughly 1 terabyte of data a day, but shipping data from Antarctica is difficult, pushing researchers toward local processing. Nvidia said researchers are using compact DGX Spark systems on site to reduce and analyze detector output locally and to test new models for identifying cosmic neutrinos.

DGX Spark targets desktop HPC and AI

Nvidia introduced an HPC toolkit for DGX Spark, extending the small-form-factor system beyond AI development into simulation and scientific code execution. The company said the machine supports MPI, can be clustered into a desk-side mini supercomputer, and is intended for local development of simulation codes, scientific large language models and agent-based workflows. The message was clear: some supercomputing work is moving closer to the researcher’s desk.

New libraries for biology, microscopy and fusion

Nvidia highlighted new software for spatial biology, where tissue imaging can involve reconstructing about 1.2 billion RNA strands across roughly 700,000 cells in a sample of around 1 square centimeter. Work with Memorial Sloan Kettering Cancer Center uses GPU software to automate reconstruction in about two seconds, enabling high-throughput analysis. Other releases included tools for real-time atomic microscopy with Lawrence Berkeley National Laboratory and visualization for fusion plasma research.

Black hole imaging accelerated from months to minutes

A centerpiece was cuPhoton, a library for radio astronomy and other data-intensive science. In black hole research tied to the M87 galaxy, Nvidia said GPU acceleration can increase data loading by up to 3,700 times and correlation by up to 460 times, reducing months of computation to minutes. The work builds on the global telescope concept that combined eight radio observatories to create humanity’s first image of a black hole and now aims at more real-time analysis.

Quantum computing effort focuses on error correction

Nvidia framed quantum computing as an error-correction problem that still needs classical supercomputers. It cited a simulation run across more than 8,000 GPUs on JUPITER that tested fault-tolerant methods and showed a 10-fold reduction in operations for one logical step through a technique called magic state cultivation. Nvidia also announced work with IBM and Fermilab to speed error decoding, with the goal of cutting response times from around 300 microseconds to 10 microseconds.

AI agents move into scientific workflows

Under a new initiative called Nemotron for Science, Nvidia is offering open models and tooling for local, on-premise scientific agents. Early examples included Argonne work on protein binder design, where an agent planned hypotheses, launched simulations, analyzed results and identified two new binder sites. Nvidia also pointed to agent tools for GROMACS configuration checking and SETI Institute radio astronomy workflows, arguing that agents can automate specialized tasks that previously required expert intervention.

New hardware roadmap centers on efficiency and mixed workloads

Nvidia announced a new Rubin HPC GPU planned for 2028, designed to balance FP64 scientific computing with AI workloads and deliver roughly three times as many CUDA cores as Hopper. The company also promoted Vera CPUs for science, including a new DKRZ supercomputer due in 2026 that is expected to double simulation performance while using about one quarter of the energy of existing CPU clusters. Networking was another focus, with co-packaged optics switches pitched as more reliable for large AI and HPC systems.

CONCLUSION

Nvidia’s strategy is to make scientific computing a blended environment where simulation, AI agents, quantum research and accelerated hardware operate as one system. The stakes are high: whichever platform best unifies those tools could shape the next decade of discovery in science and engineering.

Ask a question

More from NVIDIA