Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseResearch Paper · For R&D3DNews AI→ original

Nvidia Announces Rubin HPC: Over 50,000 CUDA Cores, Nearly 2.5x the RTX 5090

Nvidia has announced Rubin HPC, an accelerator for scientific computing: climate forecasting, structural strength analysis, genome analysis and financial risk assessment. It has over 50,000 CUDA cores (individual compute units), nearly 2.5 times as many as the gaming RTX 5090. But core count alone doesn't tell you how fast the chip is: Nvidia has not yet disclosed memory or power figures. It ships in 2028, so treat it as a planning benchmark, not a reason to buy.

AI-processed from 3DNews AI; edited by Hamidun News
Nvidia Announces Rubin HPC: Over 50,000 CUDA Cores, Nearly 2.5x the RTX 5090
Source: 3DNews AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary source3DNews AI ↗
Editor-in-Chief
Key metric #1Part of the Vera Rubin platform, but aimed at HPC rather than AI training
Key metric #2Cores sit idle less often thanks to sparse data acceleration, which helps where many values are zero
Key metric #3Price, memory, FP32/FP64 performance and manufacturing partners are still unknown

Nvidia has announced Rubin HPC, a graphics processor for classic high-performance computing (HPC): climate modeling, structural strength calculations, genome analysis and financial risk assessment. It has over 50,000 CUDA cores, the name Nvidia gives to the individual compute units on its GPUs. That is nearly 2.5 times as many as the GeForce RTX 5090, the consumer card usually treated as the benchmark for a "large" gaming GPU. The key point: core count indicates scale, not speed, and Nvidia has not yet disclosed memory, power consumption or price. Availability is expected in 2028.

What Was Actually Announced

Rubin HPC will be part of the Vera Rubin platform, but according to Nvidia it is aimed less at training and running AI models than at complex scientific and engineering workloads. The company cites climate and engineering simulation, genomic analysis, financial risk assessment and big data analytics as target areas. These are tasks where the same operations repeat billions of times over huge arrays of numbers.

The comparison with the RTX 5090 is largely a marketing device. It conveys scale: the accelerator has over 50,000 cores, while the gaming card has roughly 2.5 times fewer. But real performance also depends on memory capacity and bandwidth, and on throughput in FP32 and FP64 calculations, that is, single- and double-precision number formats that underpin most scientific simulations. Those figures are not yet available.

Why Sparse Data Acceleration Matters

One of the key features is sparse data acceleration. Sparse data consists of tables and matrices in which almost all values are zero or insignificant. Picture a table of every flight schedule in the world: only a small fraction of the cells are filled. If a processor dutifully iterates through the empty cells, it wastes time.

According to Nvidia's description, the technology is meant to feed data to the compute cores more efficiently so they spend less time waiting for information. In scientific workloads this often matters more than core count: thousands of compute units are useless if data does not reach them in time.

For data exchange between multiple GPUs, the chip supports seventh-generation NVLink, a high-speed point-to-point interconnect. It is needed when a task does not fit on a single card and the chips must exchange intermediate results.

Software and AI Agents

Rubin HPC will work with Nvidia's CUDA-X libraries, collections of optimized algorithms for scientific computing and analytics. Nvidia specifically notes that AI agents will be able to use them to run simulations and test hypotheses. The workflow is straightforward: an agent formulates a hypothesis, sends a calculation to the accelerator, receives the result and decides what to check next, without constant human involvement.

What Is Still Unknown

The list of unknowns is longer than the list of facts. Nvidia has not disclosed:

  • the exact number of CUDA cores (it says only "over 50,000");
  • video memory capacity and bandwidth;
  • power consumption;
  • FP32 and FP64 performance;
  • the accelerator's price or the cost of systems built around it;
  • the names of manufacturing partners.

Therefore, comparisons with current accelerators are premature.

What This Means in Practice

Developers of scientific software and teams that build engineering or financial models should note two things. First, Nvidia is keeping a separate HPC line rather than reducing everything to AI accelerators, so investment in CUDA-X remains a sound bet. Second, working with sparse data is becoming a hardware priority: if your algorithms can exploit such structures, they will likely benefit more from the new generation.

It is too early to plan budgets around a single core count. Wait for data on memory, FP64 and power consumption. Those will show whether Rubin HPC is a real step forward for scientific computing or stays a line in a presentation.

Story

Key Questions Answered

What Was Actually Announced?

Rubin HPC will be part of the Vera Rubin platform, but according to Nvidia it is aimed less at training and running AI models than at complex scientific and engineering workloads. The company cites climate and engineering simulation, genomic analysis, financial risk assessment and big data analytics as target areas. These are tasks where the same operations repeat billions of times over huge arrays of numbers. The comparison with the RTX 5090 is largely a marketing device.

Why Sparse Data Acceleration Matters?

One of the key features is sparse data acceleration. Sparse data consists of tables and matrices in which almost all values are zero or insignificant. Picture a table of every flight schedule in the world: only a small fraction of the cells are filled. If a processor dutifully iterates through the empty cells, it wastes time. According to Nvidia's description, the technology is meant to feed data to the compute cores more efficiently so they spend less time waiting for information.

What are the key details on Software and AI Agents?

Rubin HPC will work with Nvidia's CUDA-X libraries, collections of optimized algorithms for scientific computing and analytics. Nvidia specifically notes that AI agents will be able to use them to run simulations and test hypotheses. The workflow is straightforward: an agent formulates a hypothesis, sends a calculation to the accelerator, receives the result and decides what to check next, without constant human involvement.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…