NVIDIA canceled its plans for a quad-die Rubin Ultra AI accelerator. The company opted for a more conservative dual-GPU design instead.
Research firm SemiAnalysis reports the shift stems from manufacturing execution concerns. Engineers faced significant complexity connecting and cooling four large chiplets.
The original design required 16 HBM4E memory modules. This change theoretically halves the potential performance of the original design.
NVIDIA slated the Rubin Ultra for a 2027 release. The move signals a strategic shift in NVIDIA's data center roadmap, prioritizing manufacturability over absolute performance limits.