Samsung ships industry-first 12-layer HBM4E with 3.6 TB/s bandwidth

Samsung Electronics has started shipping samples of the world's first 12-layer HBM4E memory chips to major global customers, the company announced on May 29. The move pushes the AI memory race forward at a moment when demand for high-bandwidth memory is outpacing supply across the data center industry.

What HBM4E delivers

The 12-layer HBM4E chips achieve a stable pin speed of 14 Gbps, scalable up to 16 Gbps — more than 20% faster than Samsung's HBM4. That translates to a memory bandwidth of up to 3.6 terabytes per second per stack, a figure that matters enormously for the memory-hungry attention layers inside large language models.

Capacity gets a significant bump too: the initial 12-layer configuration ships at 48 GB per stack, up more than 30% from the previous generation. Samsung plans to extend the lineup with 32 GB (8-layer) and 64 GB (16-layer) variants based on customer demand.

The engineering behind it

Samsung built HBM4E on its 6th-generation 10nm-class DRAM process (1c) and a 4nm logic base die from Samsung Foundry — the same combination used in its HBM4 production. The company says optimizing both memory and logic architectures in tandem improved energy efficiency by 16% and thermal resistance by more than 14% compared to HBM4. Better heat dissipation is critical for dense AI accelerator trays where multiple HBM stacks sit millimetres apart on a silicon interposer.

Where it fits in the AI memory market

HBM has become one of the most strategically important components in the AI supply chain. Every Nvidia H-series and B-series GPU, every AMD Instinct accelerator, and every custom AI chip from Google or Amazon depends on stacks of HBM sitting on the same package. As AI model sizes grow, so does the demand for more bandwidth and capacity per stack.

Samsung's main rivals — SK Hynix and Micron — are on parallel development tracks. SK Hynix has been the preferred HBM4 supplier for Nvidia's Blackwell GPUs, giving Samsung a commercial incentive to push HBM4E samples to customers as early as possible. Mass production timelines will depend on customer qualification cycles, which typically take several months.

Sang Joon Hwang, Executive Vice President and Head of Memory Development at Samsung, said the company will continue to drive the growth of the global AI memory market off the back of its HBM4E announcement.

What comes next

Samsung has not committed to a mass production start date for HBM4E, saying only that it will align volume ramp with customer schedules after the initial sample and optimization phase. Given the typical qualification timeline, the earliest realistic mass production window is late 2026 or early 2027.

For the AI compute industry, 3.6 TB/s per stack is a meaningful milestone — it means future GPU generations can continue scaling model complexity without hitting a memory bandwidth wall in the near term.

Source: Samsung Newsroom