Nvidia Tests Rubin Ultra with Reduced Memory Due to HBM Shortage

Nvidia is testing modified versions of its upcoming Rubin Ultra accelerator with significantly reduced memory capacity. The shift is driven by a severe shortage of high-bandwidth memory (HBM) that has impacted the entire industry. According to The Information, the company is evaluating configurations with just 192GB or 256GB of memory—far below the originally announced 1TB—and is considering a transition from HBM4E to HBM4.

At the GTC conference earlier this year, Nvidia showcased Rubin Ultra featuring four compute dies and 1TB of HBM4E memory. However, production bottlenecks and limited manufacturing capacity for HBM have forced the company to explore alternative solutions. Notably, HBM4E’s customizable logic base complicates mass production.

The HBM shortage affects not only Nvidia but the broader market. Samsung, SK hynix, and Micron have already committed their entire HBM production capacity through 2027. SK hynix’s CEO previously warned that 2027 will be the most challenging year for supply, with constraints likely extending until 2030. In response, Nvidia is actively expanding partnerships with memory manufacturers, including a $500 billion long-term deal with SK hynix.

Despite the reduced memory capacity, Nvidia’s customers do not view this as a critical issue. For them, long-term collaboration with the company and access to cutting-edge technologies remain priorities. Reports in June also suggested a potential shift from a four-die to a two-die architecture for Rubin Ultra, further reducing memory requirements.