NVIDIA's Rubin Ultra GPU has undergone a significant design revision, according to reports from Taiwanese media, which claim that Team Green wants to ensure supply chain complexities are minimal with the new generation.
NVIDIA's Rubin Ultra Will Not Feature Four Dies On One Package, But Rather On a Single Board
NVIDIA operates with a highly aggressive product cadence, and while the firm's official disclosure is to be at an annual cycle, the supply chain partners need to act a lot more quickly, which is why the timeline is shorter for them, likely at eight to ten months. With that, we have seen NVIDIA in the past not making significant changes within architectures that could bring in a massive overhead for supply chain partners, and this is why, according to a report by Taiwan's Ctee, it is reported that Rubin Ultra will shift back to a dual-die design, similar to Rubin. Team Green will focus on HBM and other areas of advancement to deliver the extra performance.
However, industry analysts believe that the core issue behind Rubin Ultra's design adjustment is not weakening demand or product downgrading, but rather a reconfiguration of the architecture from highly packaged integration to board-level assembly.
Dual-die chiplet designs were introduced with Rubin, and, interestingly, NVIDIA showed intent to scale to a four-die design with Rubin Ultra, which made it a 'chonky' package in terms of reticle size. Initially, Rubin Ultra was supposed to feature 16 HBM4 stacks with a total capacity of 1 TB, four reticle-sized GPU dies, and CoWoS-L packaging. With such a package on board, there's a risk of thermal and structural stresses leading to 'warping', resulting in subpar manufacturing yields, which is why NVIDIA might instead go back to a dual-die configuration.
The report states that with this move, compute performance won't reduce at all, since NVIDIA will indirectly ensure a 4-die design through board-level assembly. This means that the Rubin Ultra GPU will be in a 2+2 configuration on the upcoming rack board, so there won't be a reduction in HBM4 capacity or compute from Rubin Ultra dies. A single Kyber blade will have four Rubin Ultra GPU dies, but not integrated into a single complex package, which will ultimately allow the supply chain to adjust to the changes much more easily.
No revisions are expected to the end specifications of Rubin Ultra following the change in the die approach. However, there are still questions about how NVIDIA would manage the board's physical footprint with such large Rubin Ultra dies and the thermal constraints that come with the move.
About the author: Muhammad Zuhair is a hardware and technology reporter for Wccftech, specializing in the semiconductor industry and the complex interplay between technology, manufacturing, and geopolitics. His coverage focuses on the corporate strategies and technological roadmaps of industry giants like TSMC, NVIDIA, Samsung, and Intel. Zuhair's expertise lies in deconstructing complex topics such as fabrication nodes (e.g., 2nm process), the economic impact of policies like the CHIPS Act, and the strategic development of AI infrastructure from NVIDIA, AMD and Intel.
Follow Wccftech on Google to get more of our news coverage in your feeds.
