NVIDIA is currently prototyping its next-generation Feynman GPU microarchitecture, which is set to replace the Rubin and Rubin Ultra designs. The company is leveraging TSMC's 1.6nm A16 node for production, with mass manufacturing scheduled for late 2028.
The Feynman architecture will utilize backside power delivery and advanced 3D chiplets via SoIC packaging. These designs integrate CoWoS-L 2.5D technology and panel-level packaging, alongside customized HBM4E memory featuring logic-capable base dies.
This shift aims for a massive performance leap, targeting over 1,000 TB/s in total system bandwidth using COUPE optical interconnects. This doubles the 520 TB/s capacity of Rubin Ultra, significantly advancing the Blackwell NVL72 rack's 130 TB/s baseline.