Nvidia downgrading Rubin Ultra memory specs due to HBM4E shortage
Nvidia reportedly testing lower memory configs of Rubin Ultra as memory shortage bites back — designs tested include as little as 192 GB and step back to HBM4 (external)
If the reports hold, the flagship 2027 accelerator many capacity plans assume could ship with roughly a fifth of its announced 1 TB and a step back from HBM4E to HBM4, so memory, not compute, is now the binding constraint on frontier training and inference deployments.
- The Information reports Nvidia has tested at least three lower-memory Rubin Ultra prototypes at 192 GB and 256 GB using HBM4, and versions with fewer than the 16 announced stacks, versus the 1 TB HBM4E shown at GTC.
- HBM4E's draw is a customizable base logic die (Micron/TSMC partnership let customers tweak logic per workload); that complexity is blamed for the supply strain pushing the fallback to plain HBM4.
- Reports from June say Nvidia cancelled the quad-die Rubin Ultra for a dual-GPU design; even so, the tested capacities sit below the current base Rubin GPU's 288 GB HBM4. The report does not specify die count or memory type per tested capacity.
Full summary
Citing The Information, Tom's Hardware reports Nvidia is testing Rubin Ultra accelerators with far less memory than announced, including 192 GB and 256 GB configs and a switch from HBM4E to HBM4, because it may not be able to source enough HBM. At GTC Nvidia showed a compute tray with four chiplets and 1 TB of HBM4E for the Kyber NVL144 rack targeted for 2027; SemiAnalysis says the rack slipped to 2028, while Nvidia says its roadmap is intact. HBM4E's customizable base logic die (via a Micron/TSMC partnership) is blamed for straining supply, and at least three lower-memory prototypes were tested, though die counts and per-config memory types are unclear. Separate June reports say the quad-die Rubin Ultra was cancelled for a dual-die design, and even the tested capacities fall below the current base Rubin GPU's 288 GB HBM4. Industry-wide, Samsung, SK hynix and Micron have reportedly sold through HBM capacity through 2027, with SK hynix's CEO calling 2027 the worst year for a shortage expected to last to 2030. Nvidia has moved to secure supply, expanding an SK hynix partnership into a reported $500 billion relationship with a long-term memory agreement.