Nvidia reportedly testing lower memory configurations for Rubin Ultra amid HBM shortages
Reports indicate Nvidia is planning versions of its Rubin Ultra accelerator with reduced HBM memory capacities to address supply challenges for advanced memory chips.
Technology publication The Information has reported that Nvidia is considering reducing the memory configuration for its high-performance Rubin Ultra accelerators. This move is reportedly driven by shortages of High Bandwidth Memory (HBM) chips, which are crucial for AI-focused accelerators.
Nvidia CEO Jensen Huang previously showcased the Rubin Ultra compute tray, which features four compute dies and 1 terabyte (TB) of HBM4E memory. A single GPU can house up to 1024GB of memory. Reports now suggest Nvidia is testing Rubin Ultra variants with 192GB or 256GB of memory, utilizing fewer memory stacks than previously announced.
This strategy might also involve a shift from HBM4E to HBM4 memory. While HBM4E offers advanced performance and supports custom base logic, its complex production is reportedly challenging memory manufacturers, impacting their ability to keep pace with Nvidia's needs.
The availability of HBM memory is a critical bottleneck in the production of AI accelerators. The increasing demand for more powerful memory solutions is fueled by large language models and other intensive AI workloads. Nvidia's potential adjustments to the Rubin Ultra configuration highlight broader industry challenges in securing sufficient memory supply and production capacity.
Nvidia is expected to continue efforts to secure HBM chip supplies while exploring alternative solutions, such as optimizing memory configurations, to meet growing demand and mitigate the impact of potential supply disruptions.