NVIDIA Considers Downgrading Rubin Ultra AI GPU Memory Amid HBM Supply Constraints

According to sources and market research firm TrendForce, NVIDIA is evaluating a reduction in high-bandwidth memory (HBM) specifications for its upcoming Rubin Ultra AI GPU, slated for launch in 2025. Potential adjustments include downgrading the originally planned HBM4E to HBM4 or reducing the HBM stack from 12Hi to 8Hi. This move stems primarily from uncertainties surrounding the validation timelines of HBM4E among the three major DRAM manufacturers, raising concerns that they may not complete validation and mass production on schedule. By decreasing the number of DRAM die layers per stack, more HBM stacks can be assembled under constrained total DRAM supply, thereby boosting shipment volumes of the Rubin Ultra GPU. TrendForce notes that the primary upgrade focus for Rubin Ultra lies in I/O speed, making NVIDIA more inclined to opt for reducing HBM stack height. Additionally, some AI ASIC vendors are also considering lower-capacity HBM configurations to balance product performance with supply availability.