Nvidia Reportedly Testing Rubin Ultra with Reduced Memory Configurations and HBM4

Rubin Ultra Memory Configurations Under Review

Nvidia is reportedly evaluating different memory configurations for its upcoming Rubin Ultra accelerator. These new designs include versions with 192 GB or 256 GB of memory, a reduction from the initially showcased 1 TB. The company is also reportedly testing the use of HBM4 memory, rather than the more advanced HBM4E that was previously announced for the accelerator.

HBM Supply Challenges

The reported reason for these changes is a potential shortage in the supply of High Bandwidth Memory (HBM). Specifically, the complexity of HBM4E is cited as a factor straining memory manufacturers' ability to keep pace with Rubin Ultra's planned rollout. This suggests that sourcing sufficient quantities of the higher-tier memory could be a bottleneck for Nvidia's production schedule.

Impact on Accelerator Design

The original Rubin Ultra, part of Nvidia's Kyber NVL144 design, was shown with four compute chiplets and 1 TB of HBM4E. The reported testing of lower memory configurations and HBM4 indicates a potential adjustment to the accelerator's specifications. This could also be related to earlier reports suggesting Nvidia cancelled its quad-die Rubin Ultra design in favor of a dual-GPU version due to manufacturing complexities, where less memory might be more appropriate.

HBM4 vs. HBM4E

HBM4E offers a customizable base logic die, a feature that allows customers to tailor the logic die to their specific needs. This customization, while beneficial, may contribute to the manufacturing complexity and supply strain. The shift to HBM4 would mean foregoing this specific customization capability, potentially simplifying production but also altering the accelerator's capabilities.