Renderu.com
10.3 KPosts
68.3 KFollowers
35.1 MVisits
Alex Rowan

NVIDIA is preparing a GPU with more than 50,000 CUDA cores for simulation engineers

Hardware

NVIDIA has announced Rubin HPC. This is a separate graphics processor within the Vera Rubin platform, which the company is building for simulations, engineering calculations, data analytics and AI. The chip has more than 50,000 CUDA cores. According to NVIDIA, this is a record for a single GPU and roughly three times more than Hopper. Launch is expected in 2028.

The announcement was made on 16 November 2026 at the SC26 conference in Chicago. The new product was presented by Ian Buck, NVIDIA's Vice President of hyperscale and high-performance computing. For those working with CFD, finite elements and crash simulations, this is a notable signal: the company is carving out a separate GPU within the Vera Rubin platform aimed at scientific and engineering calculations.

NVIDIA is preparing a GPU with more than 50,000 CUDA cores for simulation engineers

What is known about the chip

NVIDIA has so far disclosed little technical data. The following is confirmed:

  • more than 50,000 CUDA cores on a single GPU, roughly three times more than Hopper;
  • support for a range of precisions from double-precision FP64 to four-bit FP4;
  • acceleration of work with sparse data;
  • communication between GPUs over seventh-generation NVLink.

For a simulation engineer, the key line here is the one about FP64. Double precision is needed in tasks where rounding error accumulates step by step: in long transient CFD runs, in stiff FEM systems of equations, in climate modelling. Sparse data acceleration belongs to the same area. Stiffness matrices and meshes for fluid dynamics consist almost entirely of zeros, and, as the company explains, this capability is meant to keep the cores from sitting idle while waiting for data. Seventh-generation NVLink handles exchange between GPUs, which matters when one large mesh is distributed across several accelerators.

The announcement contains no answers to practical questions. NVIDIA did not state real FP64 performance and did not clarify whether it is delivered by the chip's own units or by emulation on tensor cores. There is no comparison with the regular Rubin either. The amount and type of memory, power consumption and the make-up of the systems that will include Rubin HPC have not yet been disclosed. There was no mention of workstations or cloud instances.

Why NVIDIA links simulation with agents

The company explains the arrival of Rubin HPC through AI agents. Agents work through variants of materials, molecules and designs, and each variant has to be checked by simulation, so the number of calculations grows. Buck described the future "science stack" like this:

"The leading edge of computing and the new science stack will be simulation, as it always has been, but combined with AI and future quantum computing".

At SC26 Buck gave examples of such agents at work. An agent from Sweden's Royal Institute of Technology KTH fixes broken GROMACS configurations on its own. Over a single night on one DGX Spark it corrected 81% of test cases. NVIDIA's own agent, the Nsight AI Porting Agent, ported a routine from the WRF weather model to the GPU and sped it up by more than 5 times. This took about 150 million tokens. For those who have ported computational code to graphics cards, this is an example of how a model can take on part of such work. Alongside the GPU the company introduced Nemotron for Science. This is an open model with a chemistry dataset that can be run locally, from DGX Spark to your own supercomputer.

NVIDIA presents Vera processor measurements in OpenFOAM and other packages

In parallel, NVIDIA expanded the platform based on the Vera processor. Buck explained that agents need a CPU to execute tool calls between model steps. The German Climate Computing Centre DKRZ chose Vera for the HLRE-5 supercomputer, whose launch is expected in early 2028. Before that, systems on Vera were announced by the Texas Advanced Computing Center TACC.

The company published the test results as a separate document, the NVIDIA Vera CPU HPC Performance Brief. It collects measurements on tasks familiar to engineers: memory bandwidth in the Stream test, the WRF weather model, fluid dynamics in OpenFOAM, crash simulation in OpenRadioss, sparse linear algebra HPCG and genomics. It also contains case studies from Los Alamos National Laboratory and TACC and a description of liquid-cooled racks. An example of the wording from the document: "Vera delivers 2.4 times higher performance per watt on the OpenFOAM model". Among the other figures: more than 3.4 times more memory bandwidth per watt, twice the performance per watt in WRF, more than twice the performance and performance per watt in HPCG. In TACC's tests against a baseline x86 system, Vera proved 2.2 times faster on MuST, 5.5 times faster on PSDNS and 7.5 times faster on MILC. All of this data has been published by NVIDIA itself.

What this means for engineers and studios

Rubin HPC is more than a year away, and so far this is a claim without key figures. Judging by the description, NVIDIA is separating scientific calculations into a distinct direction within the platform rather than leaving them entirely to AI accelerators, where double precision is generally not a priority. For simulation engineers and design engineers this could deliver accelerators better suited to their tasks. But it is too early to decide whether to plan purchases or a move to the cloud around them: neither double-precision performance, nor memory, nor price, nor integration partners have been named.

For artists, VFX studios and game developers the announcement changes nothing directly. Indirect benefit may appear through simulation packages that learn to use the chip's new capabilities. Large computing centres will probably try the new product first, as is happening with the Vera processors at DKRZ and TACC. Vendors of CAE software, first and foremost CFD and FEM solvers, will most likely have to adapt their code for sparse data acceleration, and when ordinary engineering departments feel the effect will depend on how quickly that work goes.

Which figure about Rubin HPC are you waiting for from NVIDIA first of all?

  • FP64 performance
  • Memory capacity
0
Related news:
NVIDIA Open-Sources Cosmos 3 Models for Streaming World Simulation
NVIDIA books supplier capacity years in advance
Jensen Huang has left NVIDIA — what is known about his departure
NVIDIA fits photo-based 3D reconstruction into 117 million parameters
Comments:0
post date