NVIDIA Accelerates Rubin and Vera Launch, Expected to Ship in September 2025
NVIDIA is preparing to launch new generation of server chips — Rubin R100 graphics accelerators and Vera processors. According to current information, Shipment of test samples will begin as early as September 2025, which is significantly ahead of previous expectations. Rubin's architecture will use memory HBM4 with throughput up to 13 TB/s и chiplet design with CoWoS-L packaging, moving to 3 nm manufacturing standards (TSMC N3P). This is the first time that a modular structure is used GPU in NVIDIA products, and it should provide a significant performance boost.
The Vera processors will replace the Grace line and will feature 88 custom ARM cores with up to 1.8 TB/s NVLink throughput, which represents a major leap in computing power. Full Vera Rubin NVL144 kits include 144 GPU-module with performance up to 3.6 EF FP4 and 1.2 EF FP8, 75TB of fast memory and support for new NVLink6 and CX9 interfaces. According to the statement, Rubin/Vera will be part of annual release cycle — the next architecture will come out every 12 months, but Rubin comes only half a year after Blackwell Ultra.
NVIDIA is betting on energy efficiency and scalability for data centers. Rubin will consist of two GPU reticule-sized, each with 50 PF FP4 and 288 GB HBM4. The integration of new standards should reduce the load on power supply and simplify implementation in AI infrastructure. Vera and Rubin will be released in the second half of 2026, but preliminary testing will begin this fall, in preparation for large-scale deployment on cloud platforms and supercomputers.
