🔍 Read the full analysis: What 50,000-Plus CUDA Cores On Nvidia’s Rubin HPC Could Mean For Science on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Nvidia announced Rubin HPC at SC26 on Nov. 16, describing a GPU for scientific simulation and data processing with more than 50,000 CUDA cores and seventh-generation NVLink. The chip is expected in 2028, but Nvidia has not disclosed key specifications such as native FP64 performance, memory capacity, power use, price or independent benchmark results.
Nvidia announced Rubin HPC at the SC26 supercomputing conference in Chicago on Nov. 16, describing it as a GPU for scientific simulation and data processing with more than 50,000 CUDA cores on a single chip. The company says the processor is expected in 2028; it has not yet published performance benchmarks or several specifications needed to judge its value for research and engineering workloads.
In a Nov. 16 blog post, Dan Ernst, Nvidia’s senior director of supercomputing products, said Rubin HPC is part of the Vera Rubin platform and is designed to accelerate simulation and data processing across science, engineering and analytics. Nvidia calls its core count a record but gave no exact figure beyond more than 50,000.
The company also cited a sparse data acceleration capability and seventh-generation NVLink connections between GPUs. Nvidia says the design aims to keep computing units supplied with data, rather than increasing compute capacity without addressing how quickly data can reach it. The company has not explained how the sparse-data feature works or supplied system-level details.
The core count is substantially above the figures cited for other Nvidia products, but it is not a measure of overall scientific performance. Tom’s Hardware reports that Nvidia’s current Rubin GPU has 28,672 CUDA cores; Nvidia lists 21,760 for the GeForce RTX 5090. Comparing those figures with Nvidia’s lower bound for Rubin HPC gives at least 1.7 times the cores of the current Rubin GPU and at least 2.3 times the RTX 5090. Those ratios are arithmetic comparisons, not performance results, and the final Rubin HPC core count could be higher.
Nvidia Rubin HPC · SC26 · Nov 16, 2026
Where a science GPU sits in the AI research loop
An agent suggests ideas, a simulation tests them, the result guides the next step. Nvidia’s pitch: speed up the test.
The loop and one simulation (schematic)
Six steps, as Nvidia describes them
- Agent ideaAn AI agent proposes candidate materials, molecules or designsNvidia
- TestEach idea is checked in a simulation before lab work or a prototypeNvidia
- GatherThe code reaches into an immense dataset for a handful of scattered valuesNvidia
- ComputeIt calculates a result and repeats this on an enormous scaleNvidia
- Feed the coresRubin HPC: sparse data acceleration is meant to keep more than 50,000 CUDA cores supplied with dataNvidia’s claim, not yet measured
- RefineFaster results shorten the loop of hypothesis, simulation and refinementNvidia
Source: Nvidia Blog, “NVIDIA Announces Rubin HPC — A New GPU to Accelerate Agents for Science, Engineering and Analytics”, Nov 16, 2026. Schematic, not the real chip or software.
Graphic: Powered by Thorsten Meyer AI
More Cores Won’t Settle Performance
Many scientific and engineering programs repeatedly fetch small amounts of data from large datasets before performing calculations. Nvidia says this pattern appears in fields including climate modeling, genomic analysis, engineering simulation and financial risk, and can leave processors waiting for data. Its stated aim for Rubin HPC is to improve the balance between data movement and computation, including workloads such as sparse matrix operations, stencil computations, sorting and reductions.
If the design delivers on that aim, it could reduce the time researchers spend running simulations and iterating on models or designs. Nvidia connects that possibility to AI agents that propose materials, molecules or engineering designs: those proposals still need to be tested through computation. The company’s argument is that faster simulation could shorten the cycle of hypothesis, simulation and refinement. That is a company claim about the intended benefit, not a result established by public benchmarks.
The core count alone cannot show whether the chip will be faster, more energy-efficient or more cost-effective for a particular research code. Performance depends on factors including precision support, memory capacity and bandwidth, software, power limits and the way each application uses the GPU. Those details matter especially for established scientific programs that rely on double-precision (FP64) arithmetic rather than workloads that map well to other kinds of computation.
Products Worth Considering

Ledger Nano Gen5 Crypto Wallet with Touchscreen, Black

LG 27G810A-B 27-inch UltraGear 4K UHD IPS Gaming Monitor

Logitech MX Vertical Wireless Ergonomic Mouse
As an affiliate, we earn on qualifying purchases.
Scientific Computing Needs More Than Cores
Nvidia’s January 2026 technical blog, updated in August, provides context for the FP64 question. Its table lists native FP64 vector performance of 34 TFLOPS for Hopper, 40 TFLOPS for Blackwell and 33 TFLOPS for Rubin. For FP64 matrix performance, it lists 67, 150 and 200 TFLOPS respectively, but labels the Blackwell and Rubin results as peak performance using Tensor Core-based emulation algorithms. Nvidia says those figures use multiple passes over lower-precision Tensor Cores; they are not equivalent to native FP64 vector performance.
The same blog says dedicated FP64 vector performance remains important for scientific applications that are not dominated by matrix kernels. That distinction helps explain why Rubin HPC’s eventual specifications need to include more than a CUDA-core count. Nvidia has not said how the new GPU will perform on native FP64 workloads or how its data-movement design will affect particular applications.
The Rubin HPC announcement was made during SC26, held in Chicago from Nov. 15 to 20. Nvidia presented the chip in a blog post by Ernst, rather than a separate press release identified in the supplied source material. A possible multi-die layout has been raised by Tom’s Hardware as an inference from the core count, but Nvidia has not disclosed Rubin HPC’s physical layout. That remains speculation, not a confirmed design detail.
““a record number of CUDA cores — more than 50,000 on a single GPU””
— Nvidia, in its Nov. 16 blog post
Key Specifications Are Still Missing
Nvidia has not disclosed Rubin HPC’s exact CUDA-core count, native FP64 figures, memory capacity or memory bandwidth. The announcement also provides no power consumption, price, named systems, shipping schedule beyond the expected 2028 availability, or independent benchmarks. It is not yet clear how the sparse-data capability is implemented, what the GPU’s physical layout will be, or how much the results will vary across workloads.
The comparisons with current Rubin and RTX 5090 core counts are useful only as raw architectural comparisons. They do not establish equivalent gains in simulation speed, scientific accuracy, energy use or cost. Nvidia’s claims about accelerating discovery and keeping cores fed with data have not been independently tested in the material provided.
Benchmarks Will Test the 2028 Promise
Nvidia says Rubin HPC is expected to be available in 2028. Before researchers and computing centers can assess it for procurement or deployment, they will need fuller specifications and performance results for relevant workloads, including native FP64 calculations, memory-intensive codes and the sparse operations Nvidia highlights.
The next useful milestones would be disclosures about systems containing the GPU, power and memory configurations, software support and independent benchmarks. Until those arrive, the announcement establishes Nvidia’s intended direction and a headline core-count target, but not how Rubin HPC will perform in practice or what it will cost to operate.
Key Questions
What did Nvidia announce?
Nvidia announced Rubin HPC, a GPU for science, engineering and analytics workloads, with more than 50,000 CUDA cores per GPU, a sparse-data acceleration capability and seventh-generation NVLink. It is part of the Vera Rubin platform.
When is Rubin HPC expected to be available?
Nvidia says Rubin HPC is expected in 2028. The announcement does not provide a more specific release date or name systems that will use it.
Does having more than 50,000 CUDA cores mean simulations will be faster?
Not by itself. Core count does not establish application performance. Results will depend on details such as memory, data movement, precision support, power limits and software, as well as the workload. Nvidia has not released independent benchmarks.
What important specifications has Nvidia not disclosed?
The announcement does not give an exact core count, native FP64 performance, memory capacity or bandwidth, power use, price, named systems or independent test results. Nvidia also has not explained how its sparse-data feature works.
Why is Rubin HPC being pitched for scientific computing?
Nvidia says many scientific applications repeatedly access scattered data and can spend time waiting for it. The company says Rubin HPC is designed to coordinate data movement and computation for workloads such as sparse matrices, stencils, sorting and reductions. That intended advantage still needs to be tested.
Source: ThorstenMeyerAI.com
