NVIDIA moves Feynman GPU plans to TSMC A16

NVIDIA is working on prototypes for its next-generation Feynman GPU microarchitecture, planned as the successor to Rubin and Rubin Ultra designs, according to supply chain sources and people close to TSMC. The company is reportedly choosing TSMC’s A16 node and bypassing the N2 family, using the more advanced 1.6 nm A16 design. Mass production is scheduled for the second half of 2028.

Feynman is expected to rely on backside power delivery in the A16 node, along with 3D chiplets using SoIC, CoWoS-L 2.5D integration and next-generation panel-level packaging. NVIDIA will initially use SoIC 3D stacking to combine multiple chiplets in a single package, reducing travel time between chips and lowering latency.

The stacked chips are then expected to be packaged side by side with CoWoS-L or CoPoS, creating a multi-kilowatt design capable of processing dozens of PetaFLOPS. NVIDIA also plans to work with partners on customized HBM, likely HBM4E, including a specialized base die that can contain logic such as a memory controller or packet processing unit.