NVIDIA Rubin Ultra giảm dung lượng HBM xuống còn 192GB do hạn chế về nguồn cung và công suất

The Huoxing Finance report states that on August 3, the research firm SemiAnalysis released a report indicating that NVIDIA’s preview specifications for Rubin Ultra to key customers were lower than previously market expectations. The report noted that the mainline SKU of Rubin Ultra will still use HBM4, with theoretical peak FLOPs remaining unchanged, but the HBM configuration has been reduced to 8-Hi, 192GB, down from Rubin’s previous 12-Hi, 288GB. The chip-level power consumption is approximately 1800W, similar to the initial Rubin shipping version. The report highlighted that the primary upgrade focus in Rubin Ultra’s current roadmap is on scaling interconnects at the NVL576 level. One system configuration involves interconnecting eight 72-GPU racks using NPO to enable cross-rack switch connectivity. SemiAnalysis considers this configuration one of the most plausible current candidates. SemiAnalysis stated that NVIDIA may still offer a 2600W to 2800W Max-P high-power version, as well as a 1200W version targeting lower computational loads such as token decoding. The 1800W Max-Q version better suits today’s data center power-constrained environments and offers greater deployment efficiency in terms of FLOPs/Watt. SemiAnalysis believes NVIDIA’s adjustment of Rubin Ultra’s specifications is primarily due to HBM supply constraints and data center power limitations. Reducing HBM capacity and power specifications helps improve customer deployment feasibility, lower material costs, and buffer against next year’s expected HBM price increases.