Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days

Nvidia Puts More Than 50,000 CUDA Cores on a Single Science GPU, but Rubin HPC Ships in 2028

Nvidia Puts More Than 50,000 CUDA Cores on a Single Science GPU, but Rubin HPC Ships in 2028

At SC26 in Chicago, Nvidia announced Rubin HPC, a science-focused GPU with more than 50,000 CUDA cores, sparse data acceleration and seventh-generation NVLink. Nvidia disclosed no simulation speedup, and the chip is not expected until 2028.

Nvidia used its SC26 keynote in Chicago on Nov. 16 to announce Rubin HPC, a new GPU in the Vera Rubin platform built specifically for simulation, engineering and analytics rather than for training or serving chatbots. The headline specification is core count: more than 50,000 CUDA cores on a single GPU, which Nvidia describes as a record for the company.

The design rationale is about data movement, not raw arithmetic. Nvidia argues that scientific codes share one shape: an application reaches into an immense dataset, gathers a handful of scattered values, computes a result, and repeats that loop at enormous scale. Piling on compute cores without feeding them simply leaves them idle, so Rubin HPC pairs the larger core array with a sparse data acceleration capability and seventh-generation NVLink for GPU-to-GPU traffic. The workloads Nvidia names are the recurring primitives of simulation: sparse matrix operations, stencil computations, sorting and reductions.

Two limits are worth stating plainly. Nvidia did not disclose any simulation speedup figure for Rubin HPC, in contrast with the percentage claims that usually accompany its AI accelerators. And the GPU is expected to be available in 2028, roughly two years out, which makes the announcement a roadmap marker rather than a product customers can benchmark now. "For agents tackling science and engineering, faster simulation means faster discovery," said Ian Buck, Nvidia's vice president of hyperscale and high-performance computing. "Rubin HPC accelerates the simulations agents use to test ideas and guide their next steps."

The software half of the pitch is CUDA-X, the library collection Nvidia positions as the set of optimized algorithms an agent can call while it works through a problem. The company's framing is that agent-invoked simulation shortens the loop of hypothesis, simulation and refinement, moving lab work toward the most promising candidates sooner. That framing is also a defensive argument: if scientific agents become a meaningful workload, the value accrues to whoever owns both the libraries and the silicon underneath them.

Nvidia packed two adjacent announcements into the same keynote. Nemotron for Science bundles its open Nemotron 3.5 Super VL model with a new chemistry dataset, LitmusBench v0.2, containing nearly 50,000 questions for training and evaluating chemistry reasoning, plus post-training recipes. Argonne National Laboratory is wiring Nemotron into a multi-agent system for designing protein-based therapeutic candidates, and Germany's Juelich Supercomputing Centre is integrating it. Nvidia's customer examples are company-reported and not independently verified: it says Lila Sciences reached 61 percent accuracy on making quantum dots after laboratory-feedback post-training, against 54 percent for GPT-6 Astra at 10 to 20 times lower inference cost, that Terray Therapeutics recovered about 76 percent of annotated reactions on the RxnScribe test, and that an agent built by KTH Royal Institute of Technology fixed 81 percent of broken GROMACS setups in one overnight run on a single DGX Spark.

The hardware side of that package is the Vera CPU rack-scale platform: 192 Vera CPUs interconnected by Quantum-X800 InfiniBand, each CPU carrying 88 custom Olympus cores and 176 threads with SOCAMM2 LPDDR5X memory at up to 1.2TB per second, for more than 200TB per second of aggregate memory bandwidth, roughly 300TB of memory and a petabyte of storage in a liquid-cooled MGX design. Nvidia claims 1.9 times the performance and 2.4 times the performance per watt of x86 systems on OpenFOAM's motorBike Large airflow simulation, again vendor-supplied data. Germany's DKRZ has selected Vera CPUs for its HLRE-5 climate supercomputer, and racks are expected from Dell, HPE, Lenovo and Supermicro in early 2027.

What this adds up to is Nvidia spending keynote time on the flank it has historically owned, high-performance computing, at the moment agents are being sold as research assistants. The strategic logic is sound: scientific simulation is a workload where CUDA's library depth still matters and where x86 retains a foothold. The execution risk is the calendar. A 2028 GPU with no published speedup figure asks buyers to plan around a promise, while rivals with shipping silicon get to argue from measured results. Until independent benchmarks exist, Rubin HPC is a specification sheet, not a demonstrated advantage.

Comments (0)

Log in to join the discussion

Log In

No comments yet