NVIDIA Releases Rubin HPC GPU, Betting on AI Agents to Reshape Supercomputing
At SC26, NVIDIA unveiled Rubin HPC, a new-generation GPU designed specifically for scientific computing. A single GPU has more than 50,000 CUDA cores and uses seventh-generation NVLink, with launch expected in 2028. The CUDA-X libraries are being expanded in tandem: cuPhoton.xBeam cuts black hole data analysis from one month to 8 minutes, and Daviator delivers up to a 22x speedup for VASP quantum mechanics calculations. Vera CPU racks will be supplied by Dell, HPE, Lenovo and Supermicro in early 2027.
Body · AI translation
At the SC26 supercomputing conference held in Chicago on November 16, NVIDIA unveiled Rubin HPC, a new-generation GPU designed specifically for scientific computing, engineering simulation, and data analysis, and announced a series of advances spanning CUDA-X software libraries, photonic computing, and quantum computing, in an attempt to integrate GPU-accelerated computing, AI agents, and traditional high-performance computing (HPC) into a unified platform.
This launch sends an important signal: as AI agents begin to participate in scientific research—autonomously proposing hypotheses, invoking computational tools, running simulations, and adjusting their approach based on results—the computing power demands of scientific computing are also changing. NVIDIA hopes that through hardware-software co-design, AI will not only assist researchers in analyzing data but also directly participate in the scientific discovery process from posing questions to verifying results.
Among these, Rubin HPC is the core hardware product of this launch. Unlike GPUs primarily aimed at AI training and inference, this product has been specially optimized for the massive data reads, sparse matrix operations, and complex simulation tasks common in scientific and engineering computing.
During SC 26, NVIDIA also disclosed progress on a new product: the Vera CPU rack, which will go on sale in early next year.
Overall, the signal sent by SC26 is that NVIDIA aims to compete not only in the GPU market required for AI training and inference, but also in the broader market encompassing scientific computing, engineering simulation, quantum computing, and related infrastructure. As AI agents gradually shift from information-processing tools to research assistants capable of invoking specialized software and executing complex tasks, the coordination among GPU computing power, scientific computing software, and high-speed interconnect networks will become an important direction of competition in the next phase of high-performance computing.
Rubin HPC: A single GPU with more than 50,000 CUDA cores, optimized for scientific computing
NVIDIA introduced that Rubin HPC is a newly added GPU to the Vera Rubin platform, aimed specifically at scientific, engineering, and data analysis workloads. A single GPU carries more than 50,000 CUDA cores, setting a new record for the number of cores in an NVIDIA GPU, while also introducing sparse data acceleration capabilities and adopting seventh-generation NVLink interconnect technology to improve data transfer efficiency between GPUs.
NVIDIA pointed out that scientific computing differs markedly from traditional AI training tasks. Whether it is climate modeling, engineering simulation, genomic analysis, or financial risk calculation, the relevant applications often need to repeatedly extract scattered information from massive datasets, then perform computations and repeat this process continuously. Therefore, simply adding GPU compute cores is not enough to solve the problem; data transfer and processing capabilities must also be improved in tandem to prevent compute cores from sitting idle while waiting for data.
The design focus of Rubin HPC is precisely to keep data movement and computing capability in balance, improving overall computational efficiency by accelerating common scientific computing operations such as sparse matrix operations, stencil computation, sorting, and reduction.
NVIDIA particularly emphasized the changing demands brought about by AI agents. As agents can help scientists and engineers explore more candidate materials, molecular structures, and industrial design schemes, simulation computing will become an important step in validating these schemes. After an agent proposes a hypothesis, it can invoke simulation tools to test the results and then adjust the direction of the next research step based on feedback. The faster GPU simulation is, the more schemes researchers can explore.
NVIDIA stated that Rubin HPC is expected to go on sale in 2028. Through this product, the company hopes to integrate AI agents, GPU-accelerated simulation, and scientific data processing to shorten the cycle from proposing a hypothesis to verifying a discovery.
CUDA-X expands the scientific computing landscape: black hole analysis cut from one month to 8 minutes
If Rubin HPC provides hardware computing power, then the CUDA-X software libraries that NVIDIA highlighted at this event are the key to enabling AI agents to invoke specialized scientific computing tools.
At SC26, NVIDIA announced expansion progress for the CUDA-X libraries, involving tools such as cuPhoton, cuPotential, and cuSpa, covering fields including astrophysics, materials science, drug discovery, and single-cell biology.
Among these, an application in the field of astronomy is quite representative.
The Event Horizon Telescope (EHT) links radio telescopes around the world to form a virtual telescope equivalent to the diameter of the Earth, used to observe black holes. However, the relevant observations generate petabyte-scale data, and traditional processing pipelines may take months.
The cuPhoton.xBeam module introduced by NVIDIA can use GPUs to accelerate the correlation processing of radio astronomy data, shortening an analysis task that originally took more than a month to 8 minutes. This tool has already been used to analyze observation data of the black hole jets J1924 and 3C279, and has advanced real-time image reconstruction of the black hole at the center of the Messier 87 galaxy.
In nuclear fusion research, NVIDIA demonstrated the application of cuPhoton.xLens. Researchers can extract data and analyze results synchronously while a plasma simulation is running, rather than waiting until the experiment or computation is entirely finished. Combined with the Nemotron agent running on local systems, researchers can also interactively analyze simulation results and adjust the simulation direction in real time, providing reference for the design of fusion reactor target chambers and cavities.
In drug discovery, cuPotential was developed based on research from the Simons Foundation's Flatiron Institute and D. E. Shaw Research. It can speed up existing molecular simulation software by up to 3x while improving computational stability and accuracy. The tool helps generate synthetic data for training generative AI models and helps researchers find previously unknown molecular binding sites and calculate the strength of interactions between molecules.
In materials science, NVIDIA introduced the Daviator tool. This is a set of batched GPU compute kernels for the widely used Vienna Ab initio Simulation Package (VASP), which can speed up quantum mechanical calculations by up to 22x while reducing costs by up to 75%.
In biomedicine, the newly open-sourced cuSpa can accelerate cell polygon computation, supporting large-scale spatial atlases and the construction of patient and cancer models. A related application at Memorial Sloan Kettering Cancer Center achieved up to 900x speedup compared with CPUs.
NVIDIA also introduced the rapids-singlecell framework, which can accelerate single-cell data analysis. Therna Biosciences used the tool to speed up data processing by 200x; Basecamp Research used the NVIDIA GPU-accelerated HMMER library for biological sequence annotation, achieving up to 84x speedup in related benchmarks.
These cases show that NVIDIA is trying to further develop CUDA-X from a traditional GPU computing acceleration tool into a scientific computing toolset that AI agents can directly call. Agents are not only responsible for understanding research problems and planning steps, but can also call specialized algorithm libraries to complete data processing, simulation, and analysis.
Photonic Interconnects and Quantum Computing: Completing Supercomputing Infrastructure
In addition to GPUs and scientific computing software, NVIDIA is also extending its focus this time to the network interconnection of supercomputing systems, as well as the integration of quantum computing with traditional supercomputing.
On the network infrastructure side, cloud computing service provider Lambda is applying NVIDIA's co-packaged optics (CPO) technology to production environments, used to connect AI factories equipped with GB300 NVL72 systems and more than 10,000 GPUs.
NVIDIA said Lambda's tests showed that, compared with traditional pluggable optical modules, using photonic switches reduced rack-level network latency by about 4% and pod-level network latency by up to about 10%, while maintaining comparable throughput.
NVIDIA said that by packaging optical engines together with switch chips, photonic technology can achieve up to 5x energy efficiency improvement and up to 15x reliability improvement compared with traditional pluggable optical modules, with the number of lasers reduced to one quarter of traditional solutions.
For large-scale AI clusters, a network failure can cause thousands of GPUs to wait for data at the same time, reducing the utilization of the entire system. As demand for agent inference and distributed computing increases, network latency, power consumption, and reliability will also become important factors affecting computing efficiency.
On the quantum computing side, NVIDIA is incorporating quantum processing units (QPUs) into traditional supercomputing workflows by extending the Slurm workload management system.
NVIDIA revealed that over the past five years, the number of quantum processors deployed by supercomputing centers has grown nearly 7x, and currently nearly 30 supercomputing centers worldwide are deploying more than 50 quantum processors.
Since many supercomputing centers already use Slurm to manage CPU and GPU computing tasks, NVIDIA hopes that by extending Slurm's support for quantum processors, researchers can uniformly schedule classical computing resources and quantum computing resources within familiar workflows.
This Slurm update is based on previous collaboration with IBM, Pasqal, and the quantum computing community, allowing quantum processor manufacturers to connect to different supercomputing centers through a unified interface and reducing the repetitive work of developing adaptation software for each quantum processor and different supercomputing systems.
The new version also supports interfaces such as QRMI and QDMI, and has been adopted by companies including IBM, IQM, Quantinuum, and Pasqal. Slurm 26.11 is expected to be released later in November.
At the same time, NVIDIA also expanded the CUDA-Q quantum computing platform. Among them, the new TCBP-JOBS quantum error correction decoding algorithm is up to 13,000x faster than industry-standard algorithms; Fermi National Accelerator Laboratory used CUDA-Q for related research, shortening work that usually takes five months to three weeks.
NVIDIA also demonstrated a practical application combining quantum and classical computing: QunaSys used 1,024 NVIDIA Hopper GPUs to simulate more than 300 million quantum operations and achieved chemical accuracy in a tetrahydrogen atomic model; NVIDIA researchers used Germany's JUPITER supercomputer to simulate the operation of 104 qubits under real noise conditions, for testing future fault-tolerant quantum computing architectures.
CUDA Turns 20: Developer Ecosystem Becomes a Long-Term Moat
SC26 also has a symbolic milestone: NVIDIA is celebrating the 20th anniversary of CUDA's release.
In November 2006, NVIDIA officially launched CUDA and promoted this GPU programming platform to the scientific computing field at that year's Supercomputing Conference SC06. CUDA enabled researchers to use GPUs for general-purpose parallel computing through a relatively familiar programming model, expanding the application scope of GPUs from graphics processing to scientific simulation, engineering computing, and data analysis.
Over the past 20 years, CUDA has gradually become the core of NVIDIA's accelerated computing ecosystem. NVIDIA said that more than 11 million developers worldwide now use CUDA, and CUDA-X has more than 1,000 libraries.
Representative cases cited by NVIDIA include: a University of Illinois research team used CUDA to accelerate molecular dynamics simulations, helping advance research on the atomic structure of the HIV capsid; University of Toronto researchers used two NVIDIA GPUs to train AlexNet, driving the breakthrough in deep learning in 2012; and AlphaFold 2, the protein structure prediction software that won the 2024 Nobel Prize in Chemistry, also uses CUDA for GPU computing.
From CUDA's initial service to scientific simulation and engineering computing, to later becoming an important foundation for the development of deep learning and generative AI, and now to providing AI agents with scientific algorithms and specialized computing tools, NVIDIA is further extending this software ecosystem, which has lasted 20 years, into AI-driven scientific research.
Vera CPU Rack to Launch Early Next Year, Supplied by Dell, HPE, Lenovo, and Supermicro
NVIDIA also announced that the Vera CPU rack-scale platform for high-performance computing is expected to be supplied by Dell Technologies, Hewlett Packard Enterprise (HPE), Lenovo, and Supermicro starting in early 2027.
The platform integrates 192 Vera CPUs in a single rack and interconnects them through NVIDIA Quantum-X800 InfiniBand networking. It is equipped with about 300TB of memory and 1PB of storage, uses a liquid-cooled design, and has a peak total memory bandwidth of more than 200TB/s.
NVIDIA said the platform is designed to accelerate data-intensive scientific computing tasks such as climate modeling and molecular dynamics simulation. Risk Disclosure and Disclaimer
Markets carry risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this article is at the user's own risk.
AI Trends