According to TrendForce, Nvidia has revised the design of its future “Rubin Ultra” AI accelerator by abandoning a 4-die multi-chip configuration in favor of a 2-die architecture, which is more manageable in terms of packaging and costs. The firm mentions a market launch targeted for 2027.
Why does the 4-die stall on advanced stacking?
The constraint comes from the physical limits of advanced packaging. In 4 dies, the surface area of the interposer/co-packaging would increase to approximately eight times the maximum size of a photomask, a jump which mechanically degrades yields and explodes production costs. Returning to 2 dies makes it possible to contain the size of the module, improve manufacturability and secure the increase in volume.
Rubin Ultra would remain aligned with TSMC’s N3P node with CoWoS‒L type packaging. TrendForce also indicates that Nvidia is not reducing its wafer orders at TSMC, but is reallocating more capacity to Blackwell, its current generation. Demand for 3nm is soaring: AI chips would represent 5% of 3nm capacity next year, then 36% the following year, a sign of a rapid shift from advanced lines to AI.
TSMC capacity, Nvidia arbitrations
Between an overly expensive 4-die and a 3-nm capacity window under voltage, the switch to 2-dies seems pragmatic: it reduces industrial risk in the short term while leaving room to densify memory and interconnections via CoWoS-L. In the short term, Nvidia is therefore optimizing its wafers for Blackwell, while preparing Rubin Ultra without oversizing the interposer.
Source : ITHome
Similar articles
Follow Us
