Tech media The Information published a blog post on August 6, reporting that NVIDIA is considering downgrading specifications for the Rubin Ultra GPU amid HBM memory supply shortages.
Citing three people familiar with the matter, the outlet stated that over the past several weeks, NVIDIA has internally tested at least three Rubin Ultra GPU variants with reduced HBM capacities below the originally‑announced specifications.
Lowering HBM capacity will inevitably impact the overall performance of the Rubin Ultra GPU. Nevertheless, the report notes that NVIDIA is exploring alternative approaches to mitigate the performance losses stemming from reduced HBM capacity.
Earlier coverage dated August 4 from Eugene Investment & Securities of South Korea and research firm TrendForce indicated that NVIDIA is evaluating adjustments to the HBM configuration for next‑generation Rubin Ultra AI GPUs and concurrently assessing alternative solutions besides the originally‑planned 12‑Hi HBM4E.
Potential downgrade paths under evaluation by NVIDIA include switching from HBM4E to HBM4, and reducing the HBM stack from 12‑Hi to 8‑Hi. The shift away from HBM4E stems from uncertainties in validation progress among the three major DRAM manufacturers, and it remains unclear whether validation and mass production can be completed on schedule.
Reducing the number of stacked DRAM dies within each HBM stack allows more HBM stacks to be produced from the same pool of HBM DRAM dies, which helps satisfy memory requirements for a larger volume of Rubin Ultra units.
This strategy enables higher product shipments under constrained total DRAM supply. A comparable scenario previously took place with LPDDR5X SOCAMM2 for the Vera CPU.
(Reprinted from https://news.eccn.com/）