HBF：A New Storage Solution for the AI Era
With the rapid advancement of AI technology, High Bandwidth Memory (HBM) has become a critical component for AI training and inference. However, the continuous increase in GPU computing power places higher demands on storage systems. In this context, High Bandwidth Flash (HBF) has emerged and is regarded as the next important technological direction.
HBF is a new architecture based on 3D-NAND flash memory. By stacking multiple NAND chips and utilizing Through-Silicon Via (TSV) technology for interconnects, it achieves bandwidth levels close to HBM. Its core design divides the chip into multiple sub-arrays, supporting highly parallel data processing, making it particularly suitable for high-capacity, high-throughput application scenarios.
HBF and HBM maintain a complementary relationship: HBM handles latency-sensitive hot data, while HBF is responsible for storing and reading large-scale non-volatile datasets. Leveraging the cost advantages of NAND, HBF can significantly reduce the expansion costs of AI systems, offering clear benefits in large-scale model training and edge computing.
Technical Advantages
High Bandwidth & Large Capacity
3D stacking technology enables bandwidth comparable to HBM, with storage density far exceeding that of DRAM.
Non-Volatile Nature
Data retention without power loss, supporting innovative "memory-as-storage" application models.
Significant Cost Advantage
The NAND architecture offers lower costs, making it suitable for large-scale deployment.
Challenges
1
NAND flash's write endurance and latency restrict its application in frequent-write scenarios.
2
Effective coordination with GPUs, CPUs, and other hardware is necessary to ensure optimal performance.
3
The complexity of multi-layer stacking processes may lead to yield challenges, undermining cost advantages.
4
HBF must establish its position amid competition from emerging storage technologies like 3D DRAM.
In the testing and validation phase, professional institutions such as GRGTEST can provide HBF with services including reliability testing of stacked structures, system performance evaluation, and yield analysis. These efforts help identify technical risks at an early stage and accelerate product commercialization.
In conclusion, as a vital complement to HBM, HBF is poised to play a key role in AI, data centers, and edge computing. With collaborative industry advancements and continuous improvements in testing technologies, HBF is expected to become a core component of the next-generation storage architecture.