For practitioners: High-end GPU packaging choices change thermal, yield, and system-level design trade-offs that affect data-center deployment and procurement timelines. According to reporting from Tom's Hardware, Nvidia canceled a planned quad-die version of the Rubin Ultra AI accelerator and moved to a dual-die design, citing "manufacturing execution concerns" in the Tom's Hardware piece. Wccftech's coverage, which cites Taiwanese industry sources, frames the change as a shift from highly integrated CoWoS-L packaging toward board-level or rack-level assembly, with a 2+2 arrangement of dies across a Kyber blade. Tom's Hardware reports the quad-die Rubin Ultra would have used four compute chiplets and 16 HBM4E stacks, while Wccftech notes vendors expect board-level assembly to preserve total die count per server blade. The two reports offer differing takes on performance impact: Tom's Hardware suggests the scaled-back package would be roughly half as powerful, while Wccftech reports industry sources expect compute and HBM capacity to remain intact via board-level arrangements.
Optical Scale Up Fabrics Are Limited By Manufacturing, Not Architecture