Samsung HBM4E 12-Layer 48GB — 3.6 TB/s AI Memory Stack

Specifications

Overview

HBM4E is the fifth-generation High Bandwidth Memory step, following HBM4. Samsung shipped the industry's first 12-layer HBM4E samples on 29 May 2026, roughly three weeks ahead of SK hynix's own HBM4E sample deliveries.

The headline numbers are 48 GB per 12-layer stack and 3.6 TB/s of per-stack bandwidth — a strapping of the HBM4 interface that takes per-pin signalling from HBM4's 11.7 Gbps qualified speed to a stable 14 Gbps that scales to 16 Gbps under load. That is more than a 20% per-pin improvement, and because the stack height stays at 12 layers the capacity gain over HBM4 comes from die density rather than additional stacking.

For accelerator vendors the practical consequence is that HBM4E qualification work can begin now on samples while HBM4 volume ramps, with HBM4E positioned as the dominant high-performance memory from the second half of 2026 into 2027.

Key Benefits

3.6 TB/s per stack — the highest published per-stack bandwidth in the HBM family. 12-layer package — same stack height as HBM4, so carrier and interposer designs carry over. Early samples — silicon available since May 2026 for qualification programs.

Applications

Next-generation AI training accelerators, hyperscaler custom AI ASICs, large-context inference systems, and HPC platforms that are designing against the HBM4E interface generation.

Related Products