Rubin (microarchitecture)
| Launching | 2H 2026[1] |
|---|---|
| Designed by | Nvidia |
| Manufactured by | |
| Fabrication process | 3NP or 3PN |
| Specifications | |
| Memory support | HBM4 |
| History | |
| Predecessor | Blackwell |
| Successor | Rubin Ultra (improved version) Feynman (direct successor) |
Rubin is a microarchitecture for graphics processing units (GPUs) by Nvidia.
Microarchitecture
[edit source]Announced at Computex in Taipei in 2024 by CEO Jensen Huang, it is named after the astrophysicist Vera Rubin and will consist of a GPU named Rubin and a CPU named Vera. The chips will be manufactured by TSMC using a 3 nm process and will use HBM4 memory. It is scheduled for release in Q3 of 2026.[2] Nvidia is using Blackwell GPUs to accelerate the design of Vera, Rubin, and Rubin's successor, Feynman.[3]
Rubin is said to have 50 sparse petaflops performance in FP4 (4-bit floating point math, often used for AI), increased from 20 petaflops in Blackwell, while Rubin Ultra will double the performance of Rubin with 100 sparse petaflops.[4]
Rubin Ultra
[edit source]At Nvidia GTC 2025 it was announced that Rubin will be followed by an improved Rubin Ultra architecture in 2027.[5] It would be in effect two of the Rubin cores connected together.[4]
Power delivery
[edit source]Rubin-generation accelerators are expected to draw currents in excess of 1,500 amperes at core voltages below 1 volt, which places the power delivery network (PDN) parasitic resistance on the same order of magnitude as the load impedance and makes topology choice an architectural constraint rather than a component choice. The industry's incumbent multi-phase buck point-of-load architecture, delivered by vendors such as Monolithic Power Systems, Infineon Technologies and Analog Devices (following its 2026 acquisition of Empower Semiconductor), places 12 to 24 parallel converter phases at the die perimeter and typically achieves PDN resistances of 15 to 25 microohms.[6]
Competing on-package power-delivery approaches that target the Rubin socket include factorised power modules from Vicor Corporation, which pair a pre-regulator with a Sine-Amplitude-Converter–based Voltage Transformation Module (VTM) mounted on the substrate underside, and chip-scale integrated voltage regulators from Empower (Crescendo), Ferric (Fe1766) and PowerLattice (Rainier chiplet).[7] Public commercial disclosures on Rubin socket vendor selection are expected around Q1 2027.[6]
References
[edit source]- ↑ "NVIDIA Kicks Off the Next Generation of AI With Rubin — Six New Chips, One Incredible AI Supercomputer". NVIDIA Newsroom. NVIDIA. Retrieved 5 January 2026.
- ↑ Shilov, Anton (2025-11-20). "Nvidia hints at early Vera Rubin launch — on track for $500 billion in GPU sales by late 2026 despite losing China". Tom's Hardware. Retrieved 2025-12-15.
- ↑ Nick Flaherty (March 18, 2025). "Nvidia accelerates Feynman chip design, manufacture on Blackwell GPU". eenews.
- 1 2 Sean Hollister (March 18, 2025). "Nvidia announces Blackwell Ultra GB300 and Vera Rubin, its next AI 'superchips'". The Verge.
- ↑ Jarred Walton (March 20, 2025). "Nvidia shows off Rubin Ultra with 600,000-Watt Kyber racks and infrastructure, coming in 2027". Tom's Hardware.
- 1 2 Kumar, Aditya (8 September 2026). "Vicor direct-to-pin power for AI accelerators". adikumar.co. Retrieved 9 September 2026.
- ↑ Kumar, Aditya (3 September 2026). "What Ferric is worth after ADI paid $1.5 billion for Empower". adikumar.co. Retrieved 9 September 2026.