On November 16, 2026, NVIDIA announced Rubin HPC at SC26, the international supercomputing conference. The GPU is designed for scientific and engineering simulation and large-scale data analysis. It belongs to the Vera Rubin platform and is planned to have more than 50,000 CUDA cores per GPU. NVIDIA is combining those cores with a sparse-data acceleration feature that speeds access to data scattered across memory, and with seventh-generation NVLink for connecting GPUs. It expects to deliver the product in 2028.

Scientific computing often involves extracting the needed values from huge datasets, computing on them, and repeating the process. Rather than simply adding more compute units, NVIDIA plans to strengthen how data is supplied to those units and how it moves between GPUs. At SC26, held in Chicago from November 15 to 20, the company also announced an expansion of the CUDA-X libraries used by AI agents. The aim is to make scientific research and product development more efficient: GPUs run simulations, and AI agents use the results to choose the next candidates to try.

AD

Three technologies to put 50,000+ CUDA cores to work

Rubin HPC's features can be summarized in three parts: CUDA cores that perform the computation, sparse-data acceleration that supplies the needed data to the compute units, and seventh-generation NVLink that handles communication among multiple GPUs. The main specifications shown in NVIDIA's November 16 announcement are as follows.

Role Target Announced spec
Computation Compute units of a single GPU 50,000+ CUDA cores
Data supply Access to data scattered across memory Sparse-data acceleration
GPU-to-GPU communication Data transfer between multiple GPUs 7th-generation NVLink

NVIDIA has not disclosed the internal design of the sparse-data acceleration feature or any specific performance gains.

The CUDA core count reflects the scale of the compute units, while NVLink concerns data transfer between GPUs, so the two play different roles. What NVIDIA emphasizes is reducing the time compute units spend waiting for data to arrive, leaving more time for actual computation. In addition to the performance of the compute units themselves, the company is strengthening the mechanisms that deliver data to each core and the communication between GPUs so that computing power is used efficiently. Actual performance, however, will vary with the workload, data layout, numerical precision, and GPU-to-GPU communication conditions, so comparisons require matching these conditions.

NVIDIA cites a challenge common to climate modeling, engineering simulation, genome analysis, and financial risk calculation: gathering the needed values from large datasets and computing on them repeatedly. Even when the overall dataset is enormous, the values referenced in each calculation may sit in scattered locations in memory. In such workloads, adding CUDA cores does not fully translate into higher performance if the wait for the necessary data is long.

Specific targets include sparse matrix operations, stencil computations, sorting, and reductions. A stencil computation updates each point by referencing the values of surrounding grid points, and is used in fluid analysis and weather simulation, among other fields. A reduction combines many values into a single result, for example by summing them. In these computations, reading the required data and passing along results largely determine performance, so NVIDIA says it will strengthen both computing power and data transfer performance.

Combining AI agents with scientific simulation to speed up hypothesis testing

Alongside Rubin HPC, NVIDIA is advancing the use of AI agents in scientific simulation. An AI agent that searches for candidate materials, molecules, or engineering designs runs a simulation, evaluates the result, and uses that information to choose the next hypothesis or design change. NVIDIA explains that by speeding up simulation, researchers and engineers can narrow down the next candidates to try more quickly, making experiments and design work more efficient.

Supporting this is CUDA-X, a collection of GPU-optimized algorithms and libraries. GPUs run the computation, and CUDA-X libraries handle the simulation and analysis processing that AI agents call. In its CUDA-X library expansion announced the same day, NVIDIA presented ways to connect computational results to the next research decision, using molecular simulation and fusion research as examples.

"cuPotential," for molecular simulation, is a library for evaluating candidates such as drugs computationally before they move on to actual experiments. According to NVIDIA, it speeds up existing simulation software by up to 3x while also improving the stability and accuracy of the calculations. Envisioned uses include calculating the strength of interactions between molecules and searching for the sites where molecules bind. Faster processing of this kind could shorten the time needed to evaluate candidates repeatedly and decide what to try next. However, details of the comparison environment and measurement conditions behind the "up to 3x" figure have not been released.

In the case of Pacific Fusion, a fusion research company, a method was presented for analyzing results during a computation rather than waiting for the simulation to finish. Its researchers use "cuPhoton.xLens" to extract and visualize data while a plasma simulation is running. NVIDIA explains that this entire process can be connected to a local Nemotron-based AI agent running on the same system. The idea is to analyze information obtained mid-computation interactively and reflect the results in the next simulation, informing design decisions for fusion reactor targets and chambers.

While cuPotential speeds up individual simulations, the Pacific Fusion example aims to shorten the time from in-computation analysis to the next trial. Both are examples of CUDA-X in use, and neither is a result obtained on Rubin HPC, which is scheduled for delivery in 2028. By combining such software technologies with the computing power of its next-generation GPUs, NVIDIA aims to speed up the whole sequence of forming hypotheses, running simulations, evaluating results, and revising designs.

AD

With a 2028 launch planned, real application performance is the focus

Rubin HPC is expected to be delivered in 2028, but many of the specifications needed for purchasing decisions remain undisclosed. These include double-precision floating-point (FP64) performance, which is important in scientific computing, as well as memory capacity and bandwidth, power consumption, price, and specific system configurations. For the research institutions and companies that would use it, what matters is not only whether it can handle the required precision and data scale, but also how much it can speed up existing applications.

When combined with AI agents in particular, what should be evaluated is not just the raw computing speed of a single GPU. The processing time that matters includes reading the necessary data, communicating between GPUs, analyzing simulation results, and choosing the next trial based on those results. How effective NVIDIA's 50,000+ CUDA cores and enhanced data transfer technologies prove to be will be judged by how much they can shorten the whole hypothesis-testing workflow in real scientific computing applications.