HBM3E
SK hynix HBM3E 12-Hi 36GB
36 GB twelve-high HBM3E stack for flagship AI training accelerators
SK hynix · HBM stack (2.5D) · Image: skhynix.com
12-high HBM3E delivering 36 GB per stack at 9.6 Gbps per pin — over 1.2 TB/s of bandwidth for AI training silicon.
- Interface
- HBM3E
- Form factor
- HBM stack (2.5D)
- Capacities
- 36 GB
In the High Bandwidth Memory portfolio
The 12-high HBM3E stack places 36 GB of DRAM in a single package operating at up to 9.6 Gbps per pin, over 1.2 TB/s of bandwidth per stack. An eight-stack accelerator configuration reaches 288 GB of HBM capacity — the memory foundation of current flagship AI training GPUs.
SK hynix's advanced MR-MUF packaging bonds twelve thinned die while improving heat dissipation, a decisive factor at these power densities. The company is the lead HBM supplier to the highest-volume AI accelerator programs, and this stack represents its current mass-production flagship.
Verified specification
Figures as published by SK hynix. Availability and firmware options are confirmed at quotation.
| Interface | HBM3E, 1024-bit |
|---|---|
| Stack Height | 12-high |
| Capacity | 36 GB per stack |
| Data Rate | Up to 9.6 Gbps/pin |
| Bandwidth | 1.2+ TB/s per stack |
| Packaging | Advanced MR-MUF, TSV |
| Target | AI training accelerators, HPC |
