Samsung Electronics installs HBM on top of the GPU… unveils next-generation ‘mixed-use complex semiconductor’ - 경향신문
This article was translated by an AI tool. Feedback Here.
A mockup of the next-generation 3D memory architecture ‘zHBM’, unveiled for the first time in the industry by Samsung Electronics at ‘FMS 2026’ in the United States. Provided by Samsung Electronics
Samsung Electronics unveiled ‘zHBM’, which significantly improves data processing speed and power efficiency over high-bandwidth memory (HBM), a core element of artificial intelligence (AI) infrastructure. It also introduced the high-performance NAND flash technology ‘zNAND-O’, for which demand is rising in the era of AI inference. The company plans to lead the market for next-generation memory design technologies.
On the 4th (local time), Kyung-ryun Kim, senior vice president of the DRAM Development Office at Samsung Electronics, unveiled zHBM during a keynote at the memory event ‘FMS 2026’ held at the Santa Clara Convention Center in California, United States.
Typically, HBM places memory in parallel next to the AI accelerator, but zHBM is a three-dimensional (3D) structure that stacks memory vertically on top of the AI accelerator.
This is expected to further shorten the data travel distance between the chip and memory, increasing bandwidth and power efficiency.
According to Samsung Electronics, compared even with the next-generation eighth-generation HBM5, zHBM delivers up to eight times the performance per graphics processing unit (GPU) and up to three times the performance per watt.
Samsung also showcased next-generation NAND flash technology, whose demand is increasing along with HBM as AI advances.
Jin-yeop Lee, vice president of the Flash Development Office at Samsung Electronics, unveiled the next-generation NAND solution ‘zNAND-O’, optimized for on-device (embedded) AI environments. zNAND-O uses vNAND technology, which stacks NAND vertically, to 3D-package four- or eight-die semiconductor chips.
Samsung explained that this technology enables high capacity, high bandwidth, and fast response, allowing large-scale data processing in real time. In particular, it said zNAND-O can reduce costs to one sixth while maintaining the same throughput as existing servers.
Samsung also displayed mockups (life-size models) of zHBM and zNAND-O in the exhibition area. A representative of a global AI server company said, “It was impressive to see that memory itself is evolving beyond a data storage role toward improving compute efficiency,” according to Samsung Electronics Newsroom.
It also unveiled, for the first time in the industry, ‘V10 BV-NAND’ (tenth generation), an ultra-high-stacked next-generation NAND product with more than 400 layers. Samsung said the tenth-generation NAND increases memory density by 58% compared with the previous generation (V9), allowing more data to be stored in the same area, and improves read, write, and input/output speeds.
As agent AI spreads and the focus of AI models shifts from ‘training’ to ‘inference’, innovation in memory technology is emerging as a top priority. AI inference performance depends on how efficiently a kind of ‘temporary memory’ (KV cache) that stores previous conversation content and intermediate computation data is managed.
As the storage and transmission of massive data become more important, calls for sweeping innovation in memory design technology are intensifying.
Samsung Electronics and SK hynix are also making an all-out effort to develop high-performance, high-efficiency next-generation memory technologies to alleviate AI memory bottlenecks. SK hynix, which attended this ‘FMS 2026’ event, also announced together with US NAND flash company SanDisk the previous day the first industry standard specification for vertically stacked NAND ‘high-bandwidth flash’ (HBF).
