Rubin
Nvidia GPU microarchitecture using HBM4 memory and 3 nm process
Rubin is a graphics processing unit microarchitecture designed by Nvidia. It uses TSMC 3 nm manufacturing and HBM4 memory, aiming for high AI performance with FP4 math capabilities.
- Uses HBM4 memory for high bandwidth data transfer
- Manufactured by TSMC on a 3 nm process node
- Delivers 50 sparse petaflops in FP4 precision
- Named after astrophysicist Vera Rubin
- Followed by Rubin Ultra and Feynman architectures
What it is
Rubin is a microarchitecture for graphics processing units developed by Nvidia. It targets artificial intelligence workloads and high-performance computing. The design integrates a GPU called Rubin and a central processing unit named Vera. This combination aims to boost data throughput and calculation speed. Nvidia announced the project in 2024, with deployment starting in late 2026. It represents a step forward from the previous Blackwell generation, focusing on advanced memory and process nodes.
How it came about
Jensen Huang, the chief executive of Nvidia, revealed the Rubin architecture at Computex in Taipei during 2024. Initial plans pointed to a third-quarter release in 2026. The chips moved into full production by May 2026. Deployment in data centres began in September of that year. During the GTC conference in 2025, Nvidia outlined future plans. They confirmed an enhanced version called Rubin Ultra would arrive in 2027. This improved model connects two Rubin cores to double performance metrics.
Inside the chip
The architecture draws its name from astrophysicist Vera Rubin. It pairs a GPU with a CPU core known as Vera. Nvidia utilized its own Blackwell GPUs to accelerate the design process for these new chips. The system supports High Bandwidth Memory generation four, or HBM4. This memory type allows faster data transfer rates essential for large models. Performance targets include 50 sparse petaflops in FP4 precision. FP4 refers to four-bit floating point mathematics, common in AI inference tasks. This marks a significant increase over the 20 petaflops offered by Blackwell.
How it was made
Taiwan Semiconductor Manufacturing Company produces the Rubin chips. They use a three-nanometre process node, specified as 3NP or 3PN in technical documentation. This small feature size allows more transistors to fit on the die. The manufacturing complexity requires advanced lithography tools. Nvidia relies on this partnership to maintain high yields and performance standards. The use of HBM4 memory also involves complex packaging techniques to stack memory dies close to the processor logic.
Where it went
Rubin targets artificial intelligence training and inference workloads. Its high FP4 performance makes it suitable for large language models. Data centres will deploy these units to handle massive computational loads. The integration of the Vera CPU helps manage data flow efficiently. Enterprises requiring rapid AI processing will likely adopt this hardware. It serves as a foundation for next-generation cloud computing services. The architecture supports the growing demand for generative AI applications across various industries.
Versions and family
The standard Rubin architecture serves as the base design. An improved variant called Rubin Ultra is scheduled for 2027. This version effectively links two Rubin cores together. It aims to deliver 100 sparse petaflops of performance. The direct successor to the Rubin line is named Feynman. Nvidia continues to iterate on its designs to meet increasing computational needs. Each variant builds upon the previous one with higher bandwidth and core counts.
Why it matters
Rubin continues Nvidia's dominance in the AI accelerator market. By moving to HBM4 and a 3 nm process, it sets new efficiency benchmarks. The naming convention honours scientific contributions, linking technology with research heritage. It pushes the industry toward lower precision math for higher throughput. Competitors must match these performance gains to remain relevant. The architecture supports the scaling of AI models that require vast amounts of data processing power.
Questions people ask about Rubin
What is the Rubin microarchitecture?
Rubin is a GPU microarchitecture by Nvidia designed for AI and high-performance computing. It features a GPU and a CPU named Vera, using HBM4 memory and a 3 nm process.
Who manufactures the Rubin chips?
TSMC manufactures the Rubin chips using their 3NP or 3PN three-nanometre process technology. Nvidia designs the architecture while TSMC handles the physical production.
How does Rubin compare to Blackwell?
Rubin offers 50 sparse petaflops in FP4, up from 20 petaflops in Blackwell. It also introduces HBM4 memory support and a new CPU core named Vera for better integration.
When was Rubin released?
Rubin entered full production in May 2026 and began deployment in September 2026. It was originally announced by Jensen Huang at Computex in Taipei in 2024.
Good reading and sources
- Nvidia shows off Rubin Ultra with 600,000-Watt Kyber racks and infrastructure, coming in 2027Tom's Hardware, Article
- Nvidia announces Blackwell Ultra GB300 and Vera Rubin, its next AI ‘superchips’The Verge, Article
- NVIDIA Kicks Off the Next Generation of AI With Rubin — Six New Chips, One Incredible AI SupercomputerNvidia, Article
The facts on this page come from the Wikipedia article "Rubin (microarchitecture)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryChip rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.