SemiAnalysis: NVIDIA Rubin Ultra Mainline Version’s HBM Could Fall to 192GB, With Data Center Power Supply and HBM Supply Emerging as Key Constraints

Research firm SemiAnalysis released a report stating that NVIDIA’s Rubin Ultra, previewed to key clients, has specs lower than previously market expectations. The report notes that the Rubin Ultra’s core SKU will still use HBM4, with its theoretical peak FLOPs unchanged, but its HBM configuration has been reduced to 8-Hi and 192GB, down from the prior Rubin’s 12-Hi and 288GB. The chip-level power consumption is approximately 1800W, similar to the initial Rubin shipment version. The report points out that the main upgrade focus of Rubin Ultra’s current roadmap is on NVL576-level large-scale interconnection. One system configuration involves 8 racks each with 72 GPUs interconnected, using NPO to enable cross-rack connections between switches. SemiAnalysis considers this configuration one of the more viable candidate schemes at present. SemiAnalysis adds that NVIDIA may still offer a high-power Max-P version ranging from 2600W to 2800W, as well as a 1200W version for lower compute loads such as token decoding. The 1800W Max-Q version is more aligned with current data centers’ power-constrained environments, offering better deployment efficiency in terms of FLOPs per watt. SemiAnalysis believes that the core reasons for NVIDIA’s adjustment of Rubin Ultra’s specs are tight HBM supply and data center power constraints. Reducing HBM capacity and power specifications helps improve customer deployment feasibility, lower material costs, and mitigate the pressure of HBM price hikes next year.