Nvidia's Rubin HPC GPU Packs 50,000 CUDA Cores, Ships 2028
At SC26 in Chicago, Nvidia unveiled Rubin HPC with 50,000+ CUDA cores, a 192-CPU Vera rack, science-focused Nemotron models, and photonics networking for Lambda's 10,000-GPU build.

Why it matters
- Rubin HPC carries more than 50,000 CUDA cores per GPU and ships in 2028.
- Nvidia's rack-scale Vera HPC configuration combines 192 Vera CPUs with Quantum-X800 InfiniBand per rack.
- CUDA-X's cuPhoton cuts astrophysics analysis from over a month to eight minutes; cuSpa runs up to 900x faster than CPUs.
- Lambda deploys Nvidia Photonics CPO switches connecting GB300 NVL72 racks with more than 10,000 GPUs.
- Nemotron for Science, built on Nemotron 3.5 Super VL, ships open models, a chemistry dataset, and post-training tools.
Nvidia announced Rubin HPC, a GPU with more than 50,000 CUDA cores on a single chip, at the Supercomputing '26 conference in Chicago. The new GPU, expected to ship in 2028, targets scientific simulations, engineering applications, and large-scale data processing rather than the AI training and inference workloads that dominate Nvidia's current GPU business.
The announcement anchored a broader set of releases covering next-generation GPU and CPU platforms, networking and interconnect technology, accelerated scientific computing software, and systems designed for large-scale AI workloads. Together, the SC26 news extends Nvidia's strategy of supplying complete AI and high-performance computing infrastructure rather than selling GPUs alone.
The rollout arrives as Network World noted that Nvidia may have been anxious to talk about technology rather than its recent CEO drama, and the substance of the announcements keeps the focus firmly on silicon, software, and systems.
What is Rubin HPC?
Rubin HPC is a variant within Nvidia's Vera Rubin platform, designed specifically for high-performance computing applications. Unlike GPUs optimized primarily for AI training and inference, Rubin HPC boosts scientific simulations, engineering workloads, and large-scale data processing.
The chip carries more than 50,000 CUDA cores on a single GPU. Nvidia has not disclosed the core count for the regular Rubin-generation GPUs, which leaves open the question of how the HPC variant compares to the mainstream part.
Rubin HPC includes several technical features aimed at scientific computing:
- Seventh-generation NVLink for accelerated GPU-to-GPU communication
- Sparse-data acceleration
- Optimizations for sparse matrix operations, stencil computations, and sorting and reduction operations
The ship date of 2028 places Rubin HPC firmly in Nvidia's next-generation roadmap, and the HPC-specific design signals that the company sees scientific computing as a distinct market with distinct requirements rather than a side effect of its AI business.
How does the Vera CPU fit in?
Nvidia also announced a rack-scale version of its Vera CPU platform built specifically for HPC. The move marks a shift in positioning. When Nvidia first introduced its own CPU, the company pitched it as an alternative to traditional server processors from Intel and AMD. With this announcement, Nvidia positions the CPU for a specific market segment instead.
The new configuration stacks 192 Vera CPUs per rack and pairs them with Quantum-X800 InfiniBand networking and SHARP in-network computing technology. The design reuses existing NVL72 rack components, including the chassis, power distribution, and cooling infrastructure.
That reuse matters for operators. Building an HPC rack around components already proven in Nvidia's AI systems reduces integration risk and gives supercomputing centers a familiar deployment path, the same chassis and cooling that ship with the NVL72 AI racks.
What is Nemotron for Science?
Beyond hardware, Nvidia announced Nemotron for Science, a collection of open AI models, scientific datasets, and post-training tools designed to help researchers build specialized AI agents.
Nemotron is a family of AI models Nvidia developed to power AI applications, particularly autonomous AI agents. The science platform builds on Nemotron 3.5 Super VL and includes:
- Scientific evaluation tools
- An open chemistry dataset
- Post-training recipes for specialized scientific applications
- Integration with CUDA-X accelerated computing libraries
The stated objective goes beyond question-answering. Nvidia says the agents built on the platform can interpret research information, generate code, execute computational tools, and assist with scientific workflows.
The release ties the software stack directly to Nvidia's hardware. Agents that execute computational tools and run CUDA-X libraries pull scientific workloads onto Nvidia GPUs, extending the company's CUDA software moat from AI into research computing.
How much faster is CUDA-X?
Nvidia announced additional capabilities for its CUDA-X software libraries, which accelerate specialized scientific applications. The performance claims are concrete:
- cuPhoton, for astrophysics and telescope data, reduces analysis time from over a month to eight minutes
- cuPotential, for molecular simulations, runs up to 3x faster than the prior generation
- Daviator, for materials science and quantum-mechanical calculations, runs up to 22x faster than the prior generation
- cuSpa, for biological and cellular analysis, runs up to 900x faster than CPUs
The cuPhoton figure is the standout: a reduction from more than a month of analysis to eight minutes represents a change in kind, not just degree, for telescope data processing. The cuSpa claim, measured against CPUs rather than prior-generation GPUs, shows the widest spread at 900x.
The broader objective, according to Nvidia, is to make GPU acceleration available throughout scientific workflows, including drug discovery, materials research, astrophysics, and biological research.
Why does co-packaged optics matter?
Nvidia also highlighted its photonics technology, particularly co-packaged optics (CPO) switches, alongside a customer deployment. Cloud provider Lambda is using Nvidia Photonics switches to connect GB300 NVL72 racks containing more than 10,000 GPUs.
CPO integrates optical connectivity closer to the switching silicon, reducing reliance on conventional electrical connections for high-bandwidth data movement. Nvidia says the intended benefits include:
- Lower networking power consumption
- Greater network reliability
- Improved scalability
- Better connectivity for increasingly large GPU clusters
The Lambda deployment, with more than 10,000 GPUs across GB300 NVL72 racks, is the proof point. Connecting racks of that scale is where electrical interconnect strains, and the deployment gives Nvidia a live reference for operators building comparable clusters.
The development matters because networking power consumption and bandwidth are becoming major constraints as AI clusters grow. GPU count is no longer the limiting factor; moving data between GPUs is. Nvidia's push into photonics addresses that bottleneck directly and pulls another layer of the AI data center, previously the domain of networking specialists, into Nvidia's portfolio.
What does the full picture show?
Taken together, the SC26 announcements cover every layer of a scientific computing system: the Rubin HPC GPU, the Vera CPU rack, the CUDA-X software libraries, the Nemotron for Science agent platform, and the Photonics networking that binds clusters together.
Each piece reinforces the others. The CUDA-X performance gains justify the HPC-optimized GPU; the agent platform runs on the accelerated libraries; the photonics switches connect the racks that hold the CPUs and GPUs. That vertical integration is the strategy: Nvidia no longer sells chips, it sells infrastructure.
For research institutions and supercomputing centers, the announcements signal that Nvidia intends to compete for scientific computing budgets with the same full-stack approach it applied to AI. For Intel and AMD, whose server CPUs Nvidia once positioned Vera against generically, the rack-scale HPC configuration sharpens the threat into a specific product. The 2028 ship date for Rubin HPC gives the industry three years to respond.
Source: networkworld.com
More from Rebecca Stone
Show full bio
Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
754 articles
Related articles
- NVIDIA's DGX Spark HPC Kit Puts a Four-Node Cluster on the Desk
- NVIDIA to invest up to $100 billion in OpenAI for 10 gigawatts of AI infrastructure
- NVIDIA Open-Sources Run:ai Under Apache 2.0
- Nexalus and RENDA Launch Nova8, a Liquid-Cooled AI Server Built for Retrofit
- ASUS Ramps Vera Rubin NVL72 Shipments as Rack Race Shifts to Execution