According to analysis from South Korean investment firm Eugene Investment & Securities shared by the source @jukan05, as well as an article published today by TrendForce, NVIDIA is evaluating reductions to the HBM configuration of next year's Rubin Ultra AI GPU and is simultaneously assessing alternatives to the originally planned 12Hi HBM4E.
NVIDIA's possible downgrade paths include switching from HBM4E to HBM4 and from 12Hi HBM to 8Hi HBM. The former is due to uncertainties in the HBM4E validation progress of 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 in each HBM stack would allow more HBM stacks to be built with the same quantity of HBM DRAM dies, thereby meeting the memory requirements of more Rubin Ultra units. This strategy would help ship more products when the total DRAM supply is limited. A similar situation has also occurred with the LPDDR5X SOCAMM2 used in Vera CPUs.
TrendForce believes that the main upgrade for Rubin Ultra lies in its I/O speed, so NVIDIA is more likely to adopt a strategy of reducing the HBM stack configuration. In addition, some AI ASIC companies are reportedly also considering lower HBM capacities.