NVIDIA has completed the development of six chips of the “revolutionary” Rubin architecture
NVIDIA CEO Jensen Huang announced during a working trip to Taiwan that the company completed the final stage of development of six chips of the new “revolutionary” Rubin architecture, created completely from scratch. They are currently being prepared for test production at TSMC facilities, WccfTech reports.
The Rubin platform includes R100 graphics processors, central processors, NVLink switches, silicon photonic processors and other elements, which indicates a complete restructuring of the company’s technical base.
All Rubin chips are manufactured using the 3nm TSMC N3P process and use the CoWoS-L package. For the first time ever, NVIDIA has used a chiplet design as well as a new 4x reticle, which is higher than the previous 3.3x on the Blackwell architecture. Rubin R100 GPUs will support HBM4 memory with higher bandwidth compared to HBM3E.
The Rubin architecture is focused on high-performance computing, in particular in artificial intelligence systems. Testing of thermal characteristics, power consumption and efficiency of inter-chip connections is already underway at TSMC factories. It is expected that mass production of Rubin chips will begin in 2026, and their advanced version Rubin Ultra will be released in 2027.
Jensen Huang emphasized that the main purpose of his visit to Taiwan was to assess the progress in implementing the Rubin architecture at TSMC’s production facilities. According to him, the new platform is “very advanced” and will become the foundation for the next generation of NVIDIA server solutions.
