As the parameter scale of generative AI models expands and context windows continuously lengthen, Agentic AI applications are rapidly emerging. Consequently, the developmental bottleneck of AI computing architectures has shifted dramatically. It has moved away from pure GPU processing power toward the absolute limits of memory capacity and bandwidth.
At this year’s Future Memory and Storage (FMS 2026) conference, SanDisk announced a comprehensive series of NAND innovations tailored specifically for AI inference. The most highly anticipated announcement was the High Bandwidth Flash (HBF) technology specification, released through the Open Compute Project (OCP) in collaboration with SK Hynix. Additionally, SanDisk co-developed the 10th generation BiCS10 3D NAND, stacked up to an astonishing 332 layers, with Kioxia.
HBF: The Crucial Missing Piece Complementing HBM
Currently, High Bandwidth Memory (HBM), which is based on DRAM, serves as the core memory technology in AI training and inference. Although HBM possesses exceptionally high transmission bandwidth, it faces severe technical bottlenecks regarding capacity scaling. Simultaneously, its cost and power consumption have become prohibitively high.
To bridge the vast hierarchical gap between HBM and traditional SSDs, SanDisk and SK Hynix officially released the High Bandwidth Flash (HBF) technology specification via the OCP. Unlike HBM, which relentlessly pursues ultimate bandwidth, HBF focuses intensely on maximizing data storage capacity per watt and per unit area.
HBF also employs die-stacking technology. However, it stacks NAND Flash, which offers lower costs and significantly larger capacities. According to the newly published specification, the HBF architecture supports 8-layer or 16-layer NAND stacks. A single module can achieve a maximum capacity of 512GB and deliver tiered bandwidth performance ranging from 0.4TB/s to 3.0TB/s.
More importantly, HBF utilizes the industry-standard Universal Chiplet Interconnect Express (UCIe) interface. This strategic choice enables flexible integration into various GPU and CPU packaging architectures.
This standardization initiative has progressed at lightning speed. Since the working group’s formation in February of this year, it took a mere six months to release the inaugural specification. It has already successfully attracted participation from Google DeepMind and the AI chip startup Tenstorrent.
BiCS10 QLC NAND: 332 Layers Drive a 60% Density Surge
In addition to the entirely new HBF architecture, SanDisk achieved a major breakthrough in traditional NAND technology. SanDisk and Kioxia jointly unveiled their 10th generation 3D NAND memory—BiCS10.
This next-generation product officially pushes the stacking layer count past the 300-layer threshold into entirely new territory, reaching a staggering 332 layers. Based on the Quad-Level Cell (QLC) architecture, BiCS10 boasts an incredible unit density of 37Gb/mm². Compared to current 8th-generation products, it delivers a massive 60% increase in data storage density.
Regarding performance and power consumption, BiCS10 introduces an innovative CMOS Directly Bonded to Array (CBA) design. This technique manufactures the CMOS logic circuits and the storage array on separate wafers before precisely bonding them together. Furthermore, through Toggle DDR6.0 and Separate Command/Address (SCA) technologies, it achieves a record-breaking transmission bandwidth of 4.8Gbps for a QLC architecture. Simultaneously, it utilizes Power Isolation Low Tap (PI-LTT) technology to significantly enhance energy efficiency, effectively addressing the severe heat dissipation and power consumption challenges common in high-density memory.
Empowering Enterprise SSDs and AI Data Cycles
To effectively manage the complex storage demands of the complete AI lifecycle, SanDisk also updated its proprietary “AI Data Cycle Framework” at the exhibition. This comprehensive framework encompasses raw data archiving, high-speed Data Lakes, and various application scenarios, including model storage and Retrieval-Augmented Generation (RAG).
Focusing specifically on the Key-Value (KV) cache workloads crucial to AI inference, SanDisk introduced enterprise-class SSDs featuring high-endurance configurations. These drives not only support the forthcoming PCIe Gen 6 standard but also achieve an E3.S design that perfectly balances ultra-high capacity with cost-effectiveness by minimizing DRAM requirements.
Support Our Threat Intelligence
If you find our technology report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
