Samsung and SK Hynix unveil zHBM and HBF (Hybrid Buffer Framework) memory at FMS 2026; HBF reaches 512GB with 3TB/s bandwidth, bridging HBM-to-SSD performance gap.
At the Santa Clara Convention Center in California on August 4th, FMS 2026, the memory semiconductor industry's premier event, drew over 2,000 attendees to its keynote presentations. Samsung Electronics and SK Hynix, Korea's dominant memory chipmakers, presented competing approaches to address the same emerging challenge: memory bottlenecks in the era of agentic AI.
As AI systems evolve from generative models to agentic AI that plans and executes tasks autonomously, the volume of tokens and contextual data requiring processing has surged dramatically. This shift has elevated memory—its capacity, bandwidth, and power consumption during data transfer—as a critical constraint, now exceeding the computational performance limits of graphics processing units (GPUs) themselves.
Samsung Electronics showcased zHBM, a next-generation 3D architecture that departs from the conventional placement of high-bandwidth memory (HBM) adjacent to AI accelerators. Instead, zHBM directly stacks memory on top of the AI chip, dramatically reducing data travel distance. According to Samsung, this approach delivers up to 8 times better performance per GPU and 3 times improved power efficiency compared to HBM5. Kim Kyung-ryun, an executive vice president at Samsung Electronics' DRAM Development Lab, stated: "Agentic AI is pushing performance requirements to extreme levels. Samsung is ready to develop proprietary zHBM solutions tailored to each customer's requirements and architecture."
Samsung also introduced zNAND-O, a mid-tier memory offering larger capacity than DRAM with faster data transfer speeds than conventional NAND, enabling rapid access to necessary information without storing all data in costly high-bandwidth memory. The company unveiled the industry's first 10th-generation V-NAND, the V10 BV-NAND, vertically stacked across over 400 layers. By separating memory cells and peripheral circuits onto different wafers before combining them, the technology achieves higher density and improved performance.
SK Hynix proposed an alternative strategy: hierarchical memory architecture that distributes AI data across HBM to solid-state drives (SSDs) based on importance and frequency of use, improving overall system efficiency. Working with SanDisk, SK Hynix unveiled the first standard specification for high-bandwidth flash (HBF), a NAND-based technology that complements capacity-constrained HBM. HBF specifications support NAND stacking in 8 or 16 layers, enabling storage of up to 512GB with data transfer speeds reaching 3TB per second, while implementing a common connection standard for flexible integration with AI chips from multiple manufacturers.
FMS has transformed significantly since its origins as the Flash Memory Summit, traditionally focused on storage solutions. Expanded in 2024 to encompass DRAM, HBM, and CXL memory technologies alongside storage, the event has emerged as a barometer for next-generation AI industry direction—a trajectory mirroring NVIDIA's GTC conference, which evolved from a niche gaming event into a globally watched AI showcase. Both Samsung Electronics and SK Hynix's exhibition halls drew sustained crowds throughout the event, reflecting the memory semiconductor market's emergence as a core axis of the AI industry.