Nvidia Reportedly Testing Lower-Spec Rubin Ultra GPUs with HBM4 to Address Shortages
Nvidia is reportedly testing lower-specification versions of its upcoming Rubin Ultra GPU, featuring 192GB or 256GB of HBM4 memory instead of the originally planned HBM4E. This adjustment is aimed at mitigating supply chain shortages of advanced HBM4E memory chips. As AI hardware demand continues to surge, memory supply bottlenecks remain a critical constraint for next-generation GPUs. Nvidia's willingness to downgrade memory specs highlights the manufacturing challenges associated with advanced HBM4E packaging and customizable base logic dies. The Rubin Ultra was originally showcased with up to 1TB of HBM4E memory across 16 stacks per GPU. The new test versions will feature fewer than 16 memory stacks and transition from HBM4E to HBM4, which lacks some of the complex customization features that have slowed down memory manufacturers.
## BACKGROUND
High Bandwidth Memory (HBM) is a 3D-stacked DRAM technology designed to provide ultra-high bandwidth for AI accelerators and high-performance computing. HBM4 is the next major standard, introducing a wider 2048-bit interface and a custom logic base die. HBM4E is an even more advanced iteration, but its complex manufacturing requirements have created production bottlenecks for memory suppliers.