With the rapid development of large-scale AI models, AI agents, and multimodal applications, AI computing systems are facing a new challenge: can storage capabilities keep pace with the increase in computing power? Recently, the global storage industry saw significant progress as SK Hynix and SanDisk jointly released the industry's first set of standards for High Bandwidth Flash (HBF), marking the beginning of a standardized development phase for next-generation AI storage technology.
This news has not only attracted attention from the semiconductor industry chain but also signifies that competition in AI infrastructure is gradually shifting from simply pursuing computing performance to a new stage of "computing + storage synergistic optimization." For server manufacturers, data center operators, and related hardware supply chain companies, the development of HBF may become an important direction for future AI system architecture upgrades.
Will the Establishment of the HBF Standard System Accelerate the Formation of an AI Storage Ecosystem?
The HBF standard specifications were developed under the HBF Technical Working Group of the Open Compute Project (OCP), with SK Hynix and SanDisk participating as major technical contributors. Meanwhile, Google and AI chip company Tensorrent have joined the alliance, providing support in technology verification, interface design, and ecosystem development.
This standard primarily targets AI inference systems and AI accelerator applications, establishing a unified technical framework that includes system interfaces, electrical specifications, chip stacking design, packaging reliability, and software read/write guidelines.
Compared to traditional storage solutions, HBF aims to address the data access pressure generated by the ever-expanding scale of AI models, allowing larger capacity, higher bandwidth storage resources to be closer to the computing core, thereby improving overall system efficiency.
Why does the AI industry need new storage technologies like HBF?
In recent years, AI applications, exemplified by large language models, have developed rapidly, with the scale of model parameters continuously increasing. From billions to hundreds of billions of parameters, AI systems are placing higher demands on high-speed data reading capabilities.
Currently, High Bandwidth Memory (HBM), with its extremely high data transfer speed, has become an important storage component for AI accelerators. However, due to its high cost and limited capacity, HBM is prone to capacity shortages when dealing with ultra-large-scale model inference.
HBF technology is a new storage tier that has emerged against this backdrop. It employs an advanced packaging method similar to HBM, stacking multiple NAND Flash chips to achieve greater storage capacity while maintaining high data transfer speeds.
According to relevant technology plans, the first-generation HBF product aims to achieve a single-stack capacity of 512GB and provide approximately 1.6TB/s read bandwidth, demonstrating significant advantages in capacity expansion and unit cost control.
What are the future application scenarios for HBF?
From an industry application perspective, HBF is more suitable for handling data tasks with high read demands. For example, during AI model inference, model weights, vector databases, and some cached data require frequent reads, and HBF can effectively reduce the data transfer pressure on traditional storage devices.
In the future, HBF is expected to be applied to:
AI servers and data center storage systems;
Large-scale language model inference platforms;
Intelligent driving data processing systems;
Cloud computing AI infrastructure;
High-performance computing (HPC) equipment.
Especially with the rapid development of AI agents, models need to handle more complex contextual information and real-time data interactions, further enhancing the importance of high-capacity, high-bandwidth storage.
Will HBF replace HBM? Industry competition still requires long-term observation.
While HBF has shown significant development potential, the industry generally believes it will not completely replace HBM, but is more likely to become a supplementary layer in AI storage architecture.
Because HBF is based on NAND Flash technology, its write and erase lifespan is limited, making it more suitable for applications with frequent reads and infrequent data updates, such as storing partial model parameters, historical cache data, and large-scale data retrieval tasks.
Future AI systems may form a more diversified storage architecture: HBM will handle low-latency, high-frequency access tasks, while HBF will handle larger-capacity data storage needs, with both working together to improve AI computing efficiency.
Conclusion: The release of the HBF standard will drive AI infrastructure into a new stage.
SK Hynix and SanDisk's promotion of the HBF standard is not only a technological upgrade, but also represents the AI industry chain exploring new storage solutions. As global AI applications continue to expand, achieving lower-cost, higher-efficiency data access will become key to future competition in AI infrastructure.
For semiconductor manufacturers, server equipment companies, and industrial supply chain companies, the development of the HBF ecosystem deserves continued attention. With the gradual improvement of standards and the continuous expansion of industry alliances, the next-generation AI storage market may usher in new growth opportunities.
