SemiAnalysis Explains Why Rubin Ultra's HBM Preview Shrank From 1TB to 192GB
- Source
- x.com
- Date
Myron Xie joins @JordanNanos on why NVDA's Rubin Ultra preview had 1TB of HBM per package vs the 192GB that ships. It comes down to supply and not performance, why 4-HI more than doubles the cubes you can harvest per wafer, why the loudest calls for less memory are coming from the labs themselves, and when the shortage finally eases.
Context
SemiAnalysis published a podcast segment with analyst Myron Xie explaining why Nvidia's Rubin Ultra chip shipped with 192GB of HBM memory per package instead of the 1TB shown in early previews. According to SemiAnalysis's reporting, the cut came from swapping taller memory stacks, which pack more memory dies on top of each other, for shorter 4-hi and 8-hi stacks that pack fewer dies per stack.
That tradeoff is about manufacturing yield, not chip performance. SemiAnalysis's earlier analysis explains that shorter stacks let a memory maker harvest more than double the usable memory cubes from each silicon wafer compared with taller stacks, because fewer stacked dies means fewer chances for a manufacturing defect to ruin the whole stack. Favoring higher-yield stacks over raw capacity per package is what eases the supply shortage pushing HBM prices up, even though it means Rubin Ultra ships with less total memory than its own preview showed.
Checking sign-in…
Loading comments…