🔍 Read the full analysis: The Science Case For Nvidia Rubin HPC’s 50,000-Plus CUDA Cores on ThorstenMeyerAI.com
Get the latest gadgets 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 November 16, describing a GPU for scientific simulation and data processing with more than 50,000 CUDA cores. The chip is expected in 2028, but Nvidia has not released key details such as FP64 performance, memory capacity, power, price or independent benchmarks.
Nvidia announced Rubin HPC at the SC26 supercomputing conference in Chicago on November 16, describing it as a GPU for scientific simulation, data processing and analytics with more than 50,000 CUDA cores. The company says the chip is expected to be available in 2028, but has not yet published performance benchmarks or several specifications needed to judge its value for high-performance computing.
The announcement appeared in a Nvidia blog post by Dan Ernst, senior director of supercomputing products. Nvidia says Rubin HPC is part of its Vera Rubin platform and combines a high core count with what it calls a “balanced design” intended to coordinate computing with data movement. It also cites a sparse-data acceleration capability and seventh-generation NVLink for communication between GPUs. The post does not explain the sparse-data feature in detail or give an exact core count beyond “more than 50,000.”
The announcement’s central case is that many scientific and engineering programs repeatedly retrieve small, scattered pieces of large datasets, then perform calculations across those values. Nvidia says this can leave processors waiting for data. It names sparse matrix operations, stencil computations, sorting and reductions as workloads the design aims to accelerate. These are the company’s descriptions of intended workloads, not independent measurements of Rubin HPC’s performance.
The core count is substantially above figures reported for other Nvidia GPUs, though cores alone do not establish how quickly a chip will run a given application. Tom’s Hardware reports that the current Rubin GPU has 28,672 CUDA cores; Nvidia lists 21,760 for the GeForce RTX 5090. Using those figures, the announced minimum of 50,000 is at least about 1.7 times the current Rubin count and 2.3 times the RTX 5090 count. Those ratios are arithmetic based on the cited counts, not Nvidia performance claims.
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
Why Simulation Throughput Matters
Scientific computing often depends on repeated simulation: researchers and engineers test possible materials, molecules, designs or scenarios computationally before committing resources to laboratory experiments or physical prototypes. If a GPU can shorten those runs for the relevant applications, teams could test more possibilities within a fixed computing budget. Nvidia frames the intended workflow as a faster cycle of hypothesis, simulation and refinement, including work proposed by AI agents and then checked through simulation.
That potential depends on more than the number of CUDA cores. Many scientific codes rely on double-precision (FP64) calculations, memory capacity and bandwidth, and the movement of data through a system. Nvidia’s earlier Vera Rubin platform material reports 33 TFLOPS of native FP64 vector performance for Rubin, compared with 34 TFLOPS for Hopper and 40 TFLOPS for Blackwell. Its reported FP64 matrix figures for Blackwell and Rubin use Tensor Core-based emulation algorithms, rather than native FP64 vector performance. Those platform figures do not establish Rubin HPC’s specifications; Nvidia has not said how its new GPU compares.
For research centers and operators planning future supercomputers, the announcement is therefore a direction and a timeline, not yet a basis for a performance-per-dollar decision.
Products Worth Considering

Bose QuietComfort Wireless Noise-Cancelling Headphones, Sandstone

Apple Watch Series 12 GPS 46mm Smartwatch

LG 27GR83Q-B 27-inch Ultragear QHD IPS Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Rubin HPC Fits Nvidia’s Platform
Rubin HPC was announced during SC26, held in Chicago from November 15 to 20, 2026. Nvidia’s announcement was a blog post rather than a separate press release identified in the company newsroom. It positions the chip within the Vera Rubin platform but describes a focus on science, engineering and analytics workloads rather than presenting it simply as another general-purpose AI accelerator.
Nvidia’s existing Vera Rubin platform uses sixth-generation NVLink, according to the company’s product material; the Rubin HPC announcement specifies seventh-generation NVLink. That identifies a generation change in the stated interconnect, but the post does not provide bandwidth, latency or system-level details. Tom’s Hardware suggested the core count might imply a multi-die arrangement resembling Rubin Ultra. That is an outlet’s inference: Nvidia has not disclosed Rubin HPC’s physical layout.
Nvidia’s broader technical discussion also highlights a distinction between native FP64 work and higher FP64 matrix throughput achieved through emulation on Tensor Cores. The company says native FP64 remains important for scientific applications that are not dominated by matrix kernels. That distinction matters when evaluating the new chip: the announced core total does not reveal its native FP64 capability or how it will perform on specific simulation codes.
““Designed to accelerate simulation and data processing for science, engineering and analytics workflows.””
— Nvidia, in its Rubin HPC announcement
Specs Still Missing From the Announcement
Nvidia has not disclosed Rubin HPC’s exact CUDA core count, native FP64 performance, memory capacity or bandwidth. The announcement also omits power consumption, price, named systems and independent benchmarks. Without those details, readers cannot compare its expected performance, efficiency or cost with existing accelerators for particular workloads.
The sparse-data acceleration capability is not explained, and the company has not described the GPU’s physical layout. Nor has Nvidia provided application results showing how much faster specific scientific simulations might run. The announced 50,000-plus core count is a hardware claim, not evidence by itself of a matching increase in performance. What Rubin HPC can deliver in practice remains unverified until more specifications and testing are available.
Awaiting Systems and Benchmarks
Nvidia says Rubin HPC is expected to be available in 2028. The next useful disclosures for potential customers will include detailed specifications, system configurations and performance results for representative scientific workloads. Those will help show whether its design improves simulations constrained by scattered data access, and how it handles applications that depend heavily on native FP64 computation.
Until Nvidia or system partners publish those details, the announcement establishes the GPU’s intended market, a core-count threshold and an expected release year. Pricing, deployment plans and real-world performance remain open, so researchers and computing centers do not yet have enough information to assess purchasing or migration decisions.
Key Questions
What did Nvidia announce?
Nvidia announced Rubin HPC, a GPU it says is designed for scientific and engineering simulation, data processing and analytics. The company says it has more than 50,000 CUDA cores and is expected in 2028.
Does more than 50,000 CUDA cores mean it will be faster?
Not by itself. The figure is a core-count claim, not an application benchmark. Performance also depends on factors such as memory behavior, FP64 capability, power and system design, none of which Nvidia fully specified in the announcement.
How does the announced core count compare with other Nvidia GPUs?
Using the figures cited in the source material, Rubin HPC’s stated minimum is at least about 1.7 times the 28,672 cores reported for the current Rubin GPU and 2.3 times the 21,760 cores Nvidia lists for the RTX 5090. These are core-count comparisons, not performance comparisons.
When will Rubin HPC be available?
Nvidia says the GPU is expected to be available in 2028. The company has not provided a more specific release date or named systems that will use it.
What key information has Nvidia not released?
The announcement does not provide an exact core count, FP64 figures, memory capacity or bandwidth, power consumption, price, named systems or independent benchmarks. It also gives limited detail about the sparse-data acceleration capability and chip layout.
Source: ThorstenMeyerAI.com
