Samsung Unveils ‘zHBM’, Memory Stacked on GPUs

Samsung Electronics took the stage at FMS 2026 — the world's largest memory and storage trade show, held over three days in California — to unveil a concept model of its next-generation zHBM memory architecture, a world first. The show ran August 4-6, and Samsung filled its booth with more than 30 products spanning DRAM, NAND and storage, all displayed inside a booth designed to resemble an AI cloud server.

zHBM breaks from the traditional approach of placing memory chips beside a GPU or AI accelerator. Instead, it stacks memory directly on top of the processor. Samsung says the design should deliver 4x to 8x the performance and more than 3x the power efficiency of HBM5, the eighth-generation standard that hasn't reached mass production yet. Thermal resistance — a measure of how much heat builds up — is expected to drop by more than half, and Samsung said it can customize the interlayer between the memory and accelerator for individual customers, adjusting capacity and performance to spec.

At the same event, Samsung also debuted the V10 BV NAND, a new flash memory chip stacked beyond 400 layers — another world first. The chip combines a technique that manufactures cells and circuits separately before bonding them vertically, wafer bonding, and a three-way split manufacturing process, boosting memory density 58% over the prior generation. Samsung also introduced zNAND-O, a structure that stores only the frequently accessed portions of large AI model data on NAND flash, which the company says can cut memory build costs to roughly one-sixth of current levels without sacrificing performance.

In the show's keynote, a Samsung DRAM development executive noted that token volumes used in AI inference have jumped more than 100-fold over the past two years amid the spread of agentic AI, making memory a key bottleneck in AI systems. Samsung Executive Vice President Lee Jin-yeop said Samsung is the only company that spans memory, logic chips, foundry and advanced packaging, and that it plans to use that breadth to offer integrated solutions across the entire AI semiconductor stack.