Bitget App
Trade smarter
Buy cryptoMarketsTradeFuturesStocksEarnInstitutionAI & More
Nvidia enhances scientific computing! Launches Rubin HPC GPU, betting on AI agents to reshape supercomputing

Nvidia enhances scientific computing! Launches Rubin HPC GPU, betting on AI agents to reshape supercomputing

华尔街见闻华尔街见闻2026/11/16 17:54
Show original
By:华尔街见闻

Rubin HPC features a single GPU equipped with over 50,000 CUDA cores, optimized for scientific computing, and is expected to launch in 2028. CUDA-X expands its scientific computing footprint, covering fields such as astrophysics, materials science, drug development, and single-cell biology, reducing black hole analysis from one month to just eight minutes. Nvidia showcased the integration of photonic interconnect and quantum computing, enhancing supercomputing infrastructure, stating that Lambda uses Nvidia’s photonic co-packaged optical technology (CPO) to achieve up to 5 times higher energy efficiency. Nvidia is also extending the Slurm workload management system to incorporate quantum processors (QPU) into traditional supercomputing workflows. Additionally, Nvidia announced that Vera CPU racks will be available early next year, with Dell, HPE, Lenovo, and Supermicro as suppliers.

NVIDIA is further extending the competition for AI infrastructure into the realm of scientific computing.

At the Supercomputing Conference SC26 held in Chicago on November 16, NVIDIA unveiled its next-generation GPU Rubin HPC, designed for scientific computing, engineering simulation, and data analytics. Around this core, NVIDIA announced a series of advances in CUDA-X software libraries, photonic computing, and quantum computing, aiming to integrate GPU-accelerated computing, AI agents, and traditional high-performance computing (HPC) into a unified platform.

This release sends an important signal: as AI agents begin to participate in scientific research—capable of autonomously formulating hypotheses, invoking computational tools, running simulations, and adjusting plans based on results—the demand for computing power in scientific computing is also changing. NVIDIA hopes that with co-designed hardware and software, AI will not only assist researchers in analyzing data but also directly participate in the entire process from formulating questions to verifying findings.

Of these, Rubin HPC is the core hardware product of this launch. Unlike GPUs mainly aimed at AI training and inference, this product is specially optimized for large-scale data read, sparse matrix computation, and complex simulation tasks in scientific and engineering computing.

Nvidia enhances scientific computing! Launches Rubin HPC GPU, betting on AI agents to reshape supercomputing image 0

During SC26, NVIDIA also revealed that its new Vera CPU rack will be available early next year.

Overall, the message from SC26 is that NVIDIA wants to compete not just in the GPU market for AI training and inference but also in the broader markets of scientific computing, engineering simulation, quantum computing, and related infrastructure. As AI agents shift from being information processing tools to research assistants capable of invoking professional software and carrying out complex tasks, the coordination among GPU computing power, scientific computing software, and high-speed interconnect networks will become a key focus of competition in the next stage of high-performance computing.

Rubin HPC: Over 50,000 CUDA Cores on a Single GPU, Optimized for Scientific Computing

NVIDIA describes Rubin HPC as a new GPU addition to the Vera Rubin platform, targeted at scientific, engineering, and data analytics workloads. Each GPU houses over 50,000 CUDA cores, setting a new record for the number of cores on a NVIDIA GPU, while introducing accelerated processing for sparse data and featuring seventh-generation NVLink interconnect technology for improved data transfer efficiency between GPUs.

Nvidia enhances scientific computing! Launches Rubin HPC GPU, betting on AI agents to reshape supercomputing image 1

NVIDIA points out there are clear differences between scientific computing and traditional AI training tasks. Whether it's climate modeling, engineering simulation, genomics analysis, or financial risk calculation, these applications often require repeatedly extracting dispersed information from massive datasets, then executing calculations and repeating the process. Simply increasing the GPU core count isn't enough; it's also essential to enhance data transfer and processing in tandem to avoid compute cores sitting idle while waiting for data.

The design focus of Rubin HPC is precisely on keeping data movement and compute capacity in balance, boosting overall efficiency by accelerating common scientific computational tasks such as sparse matrix operations, stencil computation, sorting, and reductions.

NVIDIA especially highlights the new demands brought by AI agents. As agents assist scientists and engineers in exploring more candidate materials, molecular structures, and industrial designs, simulation becomes the critical step to validate these plans. Agents propose hypotheses, invoke simulation tools to test results, and adjust research direction based on feedback. The faster the GPU-based simulation, the more solutions researchers can explore.

NVIDIA states Rubin HPC is expected to launch in 2028. Through this product, the company aims to integrate AI agents, GPU-accelerated simulation, and scientific data processing to shorten the cycle from hypothesis to validated discovery.

CUDA-X Expands Scientific Computing: Black Hole Analysis Down from a Month to 8 Minutes

If Rubin HPC provides hardware compute power, the CUDA-X software library—the highlight of NVIDIA’s showcase—is crucial for enabling AI agents to call professional scientific computing tools.

At SC26, NVIDIA announced expanded CUDA-X libraries, covering tools like cuPhoton, cuPotential, and cuSpa, which span fields such as astrophysics, materials science, drug development, and single-cell biology.

Nvidia enhances scientific computing! Launches Rubin HPC GPU, betting on AI agents to reshape supercomputing image 2

One representative application comes from astronomy.

The Event Horizon Telescope (EHT) connects radio telescopes worldwide to form a virtual telescope with Earth's diameter, used to observe black holes. The observations generate petabytes of data, which would take months to process using traditional workflows.

NVIDIA's cuPhoton.xBeam module leverages GPU acceleration for radio astronomy data correlation, reducing analysis that once took over a month to just 8 minutes. This tool has already been used for analyzing black hole jets J1924 and 3C279, and has advanced real-time image reconstruction of the M87 galaxy’s central black hole.

In fusion research, NVIDIA showcased the use of cuPhoton.xLens. Researchers can now simultaneously extract data and analysis during plasma simulation runs, without waiting for the entire experiment or calculation to finish. Combined with the locally-run Nemotron agent, researchers can interactively analyze simulation results and adjust models in real time—providing guidance on fusion reactor target chambers and vessel design.

For drug development, cuPotential—developed based on research from the Flatiron Institute at the Simons Foundation and D. E. Shaw Research—can speed up existing molecular simulation software by up to 3x, while improving computational stability and accuracy. The tool aids in generating synthetic data for training generative AI models and helps researchers discover previously unknown molecular binding sites or compute molecular interaction strengths.

In materials science, NVIDIA introduced Daviator, a set of GPU computing kernels for the widely used Vienna Ab initio Simulation Package (VASP). Daviator can increase the speed of quantum mechanical calculations by up to 22x and reduce costs by up to 75%.

In biomedicine, the newly open-sourced cuSpa accelerates polygon calculations for cells, supporting large-scale spatial mapping and construction of patient and cancer models. Applications at Memorial Sloan Kettering Cancer Center have achieved up to 900x acceleration compared to CPU processing.

NVIDIA also featured the rapids-singlecell framework, which speeds up single-cell data analysis. Therna Biosciences used this tool to increase data processing speed by 200x; Basecamp Research used the NVIDIA GPU-accelerated HMMER library for biological sequence annotation, with up to 84x speedup in benchmarks.

These examples show NVIDIA’s efforts to evolve CUDA-X from traditional GPU acceleration to a scientific computing toolkit that AI agents can invoke directly. Agents are not only responsible for understanding research problems and planning steps but also for calling specialized algorithm libraries to handle data processing, simulation, and analysis.

Photonic Interconnects and Quantum Computing: Enhancing Supercomputing Infrastructure

Beyond GPUs and scientific computing software, NVIDIA is also highlighting progress in supercomputer network interconnects and the integration of quantum computing with conventional supercomputing.

For network infrastructure, cloud service provider Lambda is deploying NVIDIA's Co-Packaged Optics (CPO) technology into production, connecting an AI factory equipped with GB300 NVL72 systems and over 10,000 GPUs.

Nvidia enhances scientific computing! Launches Rubin HPC GPU, betting on AI agents to reshape supercomputing image 3

NVIDIA says Lambda’s tests show that, compared to traditional pluggable optical modules, photonic switches reduce rack-level network latency by about 4% and pod-level latency by up to 10%, while maintaining equivalent throughput.

NVIDIA claims the photonic approach, which packages the optical engine with the switch chip, provides up to 5x improvement in energy efficiency and up to 15x in reliability over traditional pluggable optics, reducing the number of lasers needed to just a quarter of conventional designs.

For large-scale AI clusters, network failures can leave thousands of GPUs simultaneously waiting for data, reducing overall system utilization. As agent reasoning and distributed computing demands grow, network latency, power consumption, and reliability will become increasingly critical factors for compute efficiency.

In quantum computing, NVIDIA is expanding the Slurm workload management system to incorporate quantum processing units (QPUs) into traditional supercomputing workflows.

NVIDIA reports that in the past five years, the count of quantum processors deployed in supercomputing centers has grown nearly sevenfold, with nearly 30 centers worldwide now operating over 50 quantum processors.

Since many supercomputing centers already use Slurm to manage CPU and GPU computing tasks, NVIDIA aims to support quantum processors within Slurm so researchers can schedule both classical and quantum resources seamlessly within a familiar workflow.

This Slurm update builds on prior collaboration with IBM, Pasqal, and the quantum computing community, enabling quantum processor manufacturers to connect via a unified interface to different supercomputing centers—reducing redundant software adaptation for each pairing.

The new version also supports QRMI and QDMI interfaces, and is adopted by industry players such as IBM, IQM, Quantinuum, and Pasqal. Slurm 26.11 is expected to be released later in November.

Nvidia enhances scientific computing! Launches Rubin HPC GPU, betting on AI agents to reshape supercomputing image 4

Meanwhile, NVIDIA has expanded its CUDA-Q quantum computing platform. The new TCBP-JOBS quantum error correction decoding algorithm is up to 13,000 times faster than industry standards; Fermilab used CUDA-Q for related research, reducing work that used to take five months to just three weeks.

NVIDIA also showcased real-world quantum-classical hybrid applications: QunaSys used 1,024 NVIDIA Hopper GPUs to simulate over 300 million quantum operations and achieved chemical accuracy in the tetrahydrogen atom model; NVIDIA researchers used Germany’s JUPITER supercomputer to simulate the operation of 104 qubits under real noise, testing architectures for future fault-tolerant quantum computing.

CUDA Celebrates 20 Years: The Developer Ecosystem as a Long-Term Moat

This SC26 marks a symbolic milestone: NVIDIA is celebrating the 20th anniversary of CUDA’s release.

In November 2006, NVIDIA officially launched CUDA and debuted it at Supercomputing Conference SC06 for the scientific computing community. CUDA allowed researchers to use familiar programming models to invoke GPUs for general-purpose parallel computing, expanding GPUs’ application beyond graphics to scientific simulation, engineering computation, and data analysis.

Over 20 years, CUDA has become the core of NVIDIA’s accelerated computing ecosystem. NVIDIA claims over 11 million developers worldwide now use CUDA, and CUDA-X boasts more than 1,000 libraries.

Representative cases cited by NVIDIA include: the University of Illinois team accelerating molecular dynamics simulations with CUDA to advance atomic structure research on the HIV capsid; University of Toronto researchers training AlexNet on two NVIDIA GPUs, driving the breakthroughs of deep learning in 2012; and the 2024 Nobel Prize in Chemistry-winning AlphaFold 2 protein structure predictor, which also uses CUDA for GPU computation.

From serving initial scientific simulation and engineering, to underpinning the rise of deep learning and generative AI, and now providing scientific algorithms and specialized computation tools for AI agents, NVIDIA is further extending this two-decade-old software ecosystem into the AI-driven research frontier.

Vera CPU Racks Coming Early Next Year, Supplied by Dell, HPE, Lenovo, and Supermicro

NVIDIA also announced that the Vera CPU rack-level platform for high-performance computing is expected to be available from Dell Technologies, Hewlett Packard Enterprise (HPE), Lenovo, and Supermicro starting early 2027.

This platform integrates 192 Vera CPUs in a single rack, interconnected by NVIDIA Quantum-X800 InfiniBand networks, with about 300TB of memory, 1PB of storage, liquid-cooling design, and a peak total memory bandwidth above 200TB/s.

NVIDIA states the platform aims to accelerate data-intensive scientific computing tasks like climate modeling and molecular dynamics simulation.

0
0

Disclaimer: The content of this article solely reflects the author's opinion and does not represent the platform in any capacity. This article is not intended to serve as a reference for making investment decisions.

Understand the market, then trade.
Bitget offers one-stop trading for cryptocurrencies, stocks, and gold.
Trade now!

You may also like

Bernstein: OpenAI launches ChatGPT wallet, AI payment competition enters a new turning point, Visa (V.US) and Mastercard (MA.US) expected to become big winners

As AI agents evolve from information query tools to personal assistants capable of independently completing shopping and payment tasks, the global payments industry is facing new competitive opportunities.

智通财经•2026/11/16 16:13
Bernstein: OpenAI launches ChatGPT wallet, AI payment competition enters a new turning point, Visa (V.US) and Mastercard (MA.US) expected to become big winners