Nvidia tests multiple lower-memory Rubin Ultra configurations, including 192 GB options, as HBM supply shortage forces design flexibility.

Nvidia is reportedly testing variations of its upcoming Rubin Ultra accelerator with less memory due to concerns about sourcing sufficient HBM supply. According to The Information, some versions include just 192 GB of memory and use HBM4 instead of HBM4E, as originally announced. The report confirms an earlier assessment from SemiAnalysis about a potential Rubin Ultra memory downgrade.

Nvidia first unveiled Rubin Ultra at GTC earlier this year, showcasing a compute tray with four compute chiplets alongside 1 TB of HBM4E memory. The accelerator is part of Nvidia's Kyber NVL144 design, set to roll out in 2027, though SemiAnalysis reported the rack was delayed to 2028. When asked about the delay, Nvidia told Tom's Hardware that "our roadmap is intact" without clarifying whether the postponement was genuine.

According to The Information, Nvidia is testing Rubin Ultra versions with 192 GB or 256 GB of memory, as well as configurations using fewer than the 16 announced memory stacks. Most significantly, Nvidia is testing with HBM4 rather than HBM4E. While HBM4E incorporates traditional generational improvements, it uniquely features a customizable base logic die—a technology developed through Micron's partnership with TSMC to allow customers to modify the die for specific needs.

The complexity of HBM4E has reportedly strained supply, with memory manufacturers unable to keep pace with Rubin Ultra's rollout. Nvidia has tested at least three lower-memory designs, though details on these prototypes remain limited. The company has tested HBM4 alongside 192 GB and 256 GB configurations, with no disclosed information about compute dies or memory types for each capacity.

The number of dies is significant. In June, reports emerged that Nvidia cancelled its quad-die Rubin Ultra design due to manufacturing complexities. Though Nvidia has not commented, reports suggested the company would proceed with a dual-GPU Rubin Ultra instead—a shift that would make lower memory configurations more logical. Even so, the tested capacities fall below expectations, particularly given that each standard Rubin GPU currently ships with 288 GB of HBM4.

Nvidia is clearly preparing for a memory-constrained landscape shaped by long-standing supply agreements. In June, the company announced a partnership with SK hynix to develop next-generation memory technology, including HBM, LPDDR5X, and DDR5. In July, Nvidia expanded this relationship with a $500 billion strategic partnership that includes a long-term memory supply agreement. Notably, one Nvidia customer told The Information that per-GPU memory capacity is not a top priority, with the relationship with Nvidia itself taking precedence.

HBM shortages are affecting nearly every design on the market, though enterprise systems particularly reliant on HBM face heightened vulnerability. Last week, Digitimes reported that Samsung, SK hynix, and Micron have sold through their HBM capacity through 2027. Last month, SK Hynix CEO Kwak Noh-jung characterized 2027 as the "worst year" for the memory shortage, with supply constraints persisting through 2030.