Nvidia May Lower Rubin Ultra GPU Specs Due to HBM Shortage
Nvidia is reportedly testing at least three lower High Bandwidth Memory (HBM) configurations for its upcoming Rubin Ultra GPU to cope with supply shortages. The company is evaluating downgrading from HBM4E to HBM4, or reducing the stack height from 12Hi to 8Hi HBM. This potential downgrade highlights the severe constraints in the AI hardware supply chain, particularly regarding advanced memory packaging. It could impact the peak performance of Nvidia's next-generation AI accelerators, forcing clients to adjust their performance expectations. Reducing the stack height from 12Hi to 8Hi allows Nvidia to produce more HBM stacks from the same volume of DRAM dies, maximizing GPU shipments. However, the validation timeline for HBM4E among the top three DRAM manufacturers remains uncertain, prompting these contingency plans.
## BACKGROUND
High Bandwidth Memory (HBM) uses 3D-stacked DRAM dies connected vertically to provide massive memory bandwidth for AI accelerators. The "Hi" designation (like 8Hi or 12Hi) refers to the number of stacked DRAM layers, with higher stacks offering larger capacity but facing greater manufacturing and thermal challenges. A global HBM shortage, driven by the explosive demand for AI data centers, is projected to persist for several years.