11 hours ago
8
SK hynix has started shipping 12-layer HBM4E memory samples to major customers, pulling its timeline forward by several months and intensifying what’s become the most consequential hardware race in the AI era.
The samples pack 48GB of capacity and hit speeds of up to 16 Gbps per pin. Power efficiency improved by over 20% compared to earlier generations.
The specs and the tech behind them
SK hynix credits much of the improvement to its Advanced MR-MUF technology, which cuts heat resistance by 17%. Stacking 12 layers of memory on top of each other generates a lot of heat, and this packaging method keeps things cool enough to work reliably under sustained AI workloads.
The company had originally planned to introduce HBM4E samples in the second half of 2026. Faster-than-expected development pushed that timeline up to June, with shipments beginning on June 18.
Samsung got there first, barely
Samsung Electronics announced its own 12-layer HBM4E sample shipments on May 29, roughly three weeks before SK hynix. Samsung’s samples feature equivalent specifications, including bandwidth capabilities of up to 3.6 TB/s.
SK hynix currently captures approximately 57% of HBM industry revenue. Both companies are targeting 2027 for volume production.
Why Nvidia is the elephant in the room
SK hynix serves as a primary supplier for Nvidia’s AI GPUs. Both companies are now in the customer qualification phase, which will determine production allocation for 2027 and beyond.
The 48GB capacity per stack means fewer stacks needed per GPU package, or alternatively, more total memory per accelerator without increasing the physical footprint.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·