Nvidia has redesigned its Rubin Ultra GPU, moving from a single four-die package to a dual two-die configuration, according to Taiwan's Commercial Times. The change is tied to packaging constraints and inadequate yield on the original four-die design.
The Taiwanese trade paper also reports that Nvidia has redirected some Rubin wafer start allocations to support additional Blackwell production, a move that fits with suggestions made at Nvidia's GTC conference last month that the Rubin ramp is running slightly behind schedule.
A pivot back to Blackwell
The reallocation of wafer starts is the more consequential near-term development. Blackwell remains Nvidia's primary revenue driver, and any capacity freed up by pushing Rubin back will be absorbed by demand that continues to outpace supply. For customers waiting on Blackwell hardware, the additional production could offer some relief, though the overall supply picture stays tight.
The 3nm bottleneck
Commercial Times identifies limited availability of TSMC's 3nm process node as a structural constraint across the AI chip sector, not just for Nvidia. That constraint shapes the competitive dynamics of the industry as much as any individual company's design choices.
Analysts tracking the sector argue that tight TSMC wafer supply, combined with ongoing shortages of DRAM and high-bandwidth memory, will prevent any significant shift in market share from current allocations through 2026 and probably into 2027. For Nvidia, that is a form of protection: rivals cannot easily close the gap when the raw materials to do so are rationed.
What the redesign means for Rubin
Related reading
- Nvidia's next AI supercomputer has 1.3 million parts, weighs two tons and costs up to $4 million. It may also…
- Microsoft validates NVIDIA Vera Rubin NVL72 in the cloud
- Nvidia partners with energy giants to make AI data centres flex with the power grid
Switching from a four-die to a dual two-die package is not a trivial change, but it is a pragmatic one. Yield problems on large, complex die configurations are common at leading-edge nodes, and splitting the design reduces the risk of losing entire packages to defects. The trade-off is additional complexity in how the dies are connected and coordinated, though Nvidia's engineering record at this scale suggests that is a manageable problem.
The net effect is a Rubin Ultra that arrives later than originally expected but on a more reliable production footing once it does.
