Nvidia reportedly tests lower memory configurations for Rubin Ultra GPUs amid HBM shortage
Nvidia is reportedly testing lower memory configurations for its upcoming Rubin Ultra GPUs, including options as low as 192 GB and potentially stepping back to HBM4 instead of HBM4e. This shift is driven by ongoing shortages in the high-bandwidth memory (HBM) supply chain. This adjustment highlights the severe supply chain constraints in the semiconductor industry, which could force AI developers to replan their hardware procurement. It also demonstrates how memory bottlenecks continue to dictate the deployment timelines and specifications of next-generation AI accelerators. While the Rubin Ultra was originally designed to feature massive memory capacities using advanced HBM4e, the tested lower-tier configurations could significantly reduce the available bandwidth and capacity. The shortage affects major HBM manufacturers like SK Hynix, Samsung, and Micron, forcing Nvidia to explore these contingency designs.
## BACKGROUND
High Bandwidth Memory (HBM) is a specialized 3D-stacked DRAM interface used in high-performance AI accelerators to provide massive data throughput. The upcoming HBM4 standard doubles the I/O pin count to 2,048, significantly reducing memory bottlenecks compared to previous generations. Nvidia's Rubin architecture, expected to succeed the Blackwell architecture, relies heavily on these next-generation memory standards to power advanced AI training and reasoning.