NVIDIA has made significant changes to the architecture of the future Rubin Ultra graphics accelerator. According to Taiwanese media, the redesign is due to a desire to reduce supply chain complexities and increase production stability.
Despite the official strategy of annual product updates, manufacturing partners are working on a tighter schedule — approximately 8 to 10 months. In such conditions, radical architectural changes can create a serious burden on the supply chain. That is why Rubin Ultra, reportedly, will return to a two-chip configuration, similar to the base Rubin, instead of a more ambitious option.
Previously, it was assumed that Rubin Ultra would have a four-chip design using four dies, 16 stacks of HBM4 memory with a total volume of up to 1 TB, and complex CoWoS-L packaging. However, this approach increased the risks — in particular, overheating and deformation, which could negatively affect the yield of suitable chips.
Analysts emphasize that the changes are not related to a decrease in demand or a deterioration in performance. On the contrary, it is about moving from complex integration within a single package to a more flexible assembly at the board level.