SemiAnalysis: NVIDIA's Hopper Ultra Mainstream HBM Variant Could Drop to 192GB, Datacenter Power Delivery & HBM Supply Key Constraints

August 3rd. Research firm SemiAnalysis released a report stating that NVIDIA's Rubin Ultra, previewed to key customers, is below the market's previous expectations. The report indicates that the Rubin Ultra's mainstream SKU will still use HBM4, maintaining the theoretical peak FLOPs, but the HBM configuration will be reduced to 8-Hi, 192GB, lower than the previous Rubin's 12-Hi, 288GB. The chip-level power consumption is around 1800W, similar to the early shipping version of Rubin.

The report points out that the key upgrades in the current roadmap of Rubin Ultra are focused on the NVL576-level scalable interconnect. One system form is an interconnection of 8 72-GPU racks, using NPO to achieve cross-rack connections between switches. SemiAnalysis believes that this form is one of the more likely candidate solutions at the moment.

SemiAnalysis stated that NVIDIA may still offer a high-power 2600W to 2800W Max-P version, as well as a 1200W version for lower computing load tasks such as token decode. The 1800W Max-Q version is more suitable for current data center power-constrained environments, exhibiting higher deployment efficiency in terms of FLOPs/Watt.

SemiAnalysis believes that NVIDIA's adjustment of the Rubin Ultra specifications is mainly due to HBM supply constraints and data center power limitations. Lowering the HBM capacity and power consumption specifications will help improve customer deployment feasibility, reduce material costs, and mitigate the pressure of next year's HBM price increase.