SK hynix is developing hybrid bonding for HBM5 to break the 775-micron stack height ceiling, per Tom's Hardware. The firm extends its MR-MUF technology through Nvidia's Rubin platform.
Key facts
- 775-micron stack height limit for current HBM
- SK hynix extends MR-MUF through Nvidia Rubin
- Hybrid bonding planned for HBM5
- MR-MUF used for HBM3E and HBM4
- Hybrid bonding could cut layer thickness 30-40%
SK hynix is pushing hybrid bonding for HBM5 as AI memory stacks approach a 775-micron height limit, according to Tom's Hardware. The company will extend its existing MR-MUF (Mass Reflow Molded Underfill) technology through Nvidia's Rubin platform, while reserving hybrid bonding for the next-generation HBM5.
The 775-micron wall
The 775-micron figure represents the maximum stack height for current HBM designs, constrained by thermal dissipation and warpage. As stacks grow taller—HBM4 pushes to 12-16 layers—the mechanical and thermal stresses increase, making conventional solder-bump interconnects untenable. Hybrid bonding replaces solder bumps with direct copper-to-copper connections, reducing the height per layer and improving thermal conductivity.
SK hynix's strategy is a dual-track approach: MR-MUF remains the workhorse for HBM3E and HBM4, while hybrid bonding becomes the enabler for HBM5. This mirrors industry trends—Samsung and Micron are also exploring hybrid bonding, but SK hynix's extension of MR-MUF through Rubin suggests a more conservative near-term roadmap.
Why the ceiling matters
AI accelerators like Nvidia's Rubin demand higher memory bandwidth and capacity, which directly translates to taller stacks. The 775-micron ceiling is a physical bottleneck that limits how many DRAM dies can be stacked. Hybrid bonding could reduce per-layer thickness by 30-40%, allowing more dies within the same height envelope.
However, hybrid bonding is not trivial to implement at scale. It requires pristine surface flatness and particle-free environments, which raises manufacturing costs. SK hynix's decision to keep MR-MUF for Rubin suggests they are not yet confident in hybrid bonding's yield at high volume.
The company did not disclose specific HBM5 thickness targets or timeline, but the implication is clear: hybrid bonding is the answer to the 775-micron wall, and the race to HBM5 will be defined by who can master it first.
What to watch
Watch for SK hynix's HBM5 sample shipments in 2026 and whether hybrid bonding yields exceed 90% at volume. Also track Nvidia's Rubin GPU memory specifications—if Rubin requires more than 8 stacks, hybrid bonding becomes mandatory, accelerating the transition.
