Samsung has begun shipping samples of its new HBM4E high‑bandwidth memory ahead of the promised “later this year” window. The 12‑layer stack now reaches 48GB, up from 36 GB in the original HBM4, while keeping the same physical footprint. Early adopters in AI accelerators will be the first to test the higher capacity.

Performance climbs roughly 20 %, with per‑pin speed jumping to 14Gbps, delivering 3.6 TB/s per stack—identical aggregate bandwidth to HBM4 but with a 33 % capacity boost. Samsung achieved the gains using a 10 nm class “1c” memory die paired with a 4 nm logic base, cutting power draw by 16 % and thermal resistance by at least 14 %.

Samsung also plans 8‑layer 32 GB and 16‑layer 64 GB variants, giving designers flexibility as AI models grow. Executive Vice President Sang Joon Hwang framed HBM4E as a step toward accelerating the global AI memory market, noting the architecture can scale to 16 Gbps per pin later. The shipment marks the first commercial rollout of a memory tier that pushes capacity while easing cooling demands.

For system builders, the cooler operation—thanks to a 14 % drop in thermal resistance—means lower fan speeds and tighter board layouts. As data centers chase ever‑larger language models, the extra bandwidth and density of HBM4E could translate into measurable performance gains without a proportional power penalty.