NVIDIA CEO Jensen Huang announced during a working visit to Taiwan that the company has completed the final stage of development of six chips of the new "revolutionary" Rubin architecture, which is designed completely from scratch. The new chips are currently being prepared for trial production at TSMC factories. This is reported by WccfTech.
The new platform includes R100 GPUs, CPUs, NVLink switches, silicon photonic processors, and other components, indicating a complete overhaul of the technical stack.
All Rubin chips are built on TSMC's 3nm N3P process and use the CoWoS-L package. They feature a chiplet design for the first time in NVIDIA's history, as well as a new 4x aiming grid, which is an improvement over the 3.3x in the previous Blackwell architecture. The Rubin R100 GPUs will support HBM4 memory, which has higher bandwidth compared to HBM3E.
The Rubin platform is aimed at high-performance computing, particularly for artificial intelligence systems. NVIDIA is testing the thermal characteristics of the architecture, power consumption, and efficiency of inter-chip connections at TSMC factories. It is expected that mass production of Rubin chips will begin in 2026, and the expanded version, Rubin Ultra, in 2027.
Jensen Huang said the main purpose of the trip to Taiwan was to assess the progress of the Rubin architecture at TSMC. According to him, the new platform is "very advanced" and will be the basis for NVIDIA's next generation of server solutions.