SK hynix's High Bandwidth Flash (HBF), unveiled at FMS 2026, is expected to enable the attachment of a large memory pool for AI accelerators, allowing the same GPU to handle more queries while reducing power and total cost of ownership.
According to the semiconductor industry on August 6, SK hynix's HBF strategy is seen as an attempt to alleviate bottlenecks in AI inference infrastructure by rearranging memory layers. This approach introduces a new layer between HBM and SSD, moving beyond the competition to enhance HBM performance.
HBF is based on NAND technology but differs from standard SSDs. It is designed with NAND in a stacked package similar to HBM, enhancing parallel reading structures to bring it closer to AI accelerators. The specifications released by SK hynix and SanDisk indicate a maximum capacity of 512GB and a bandwidth ranging from 0.4 to 3.0TB/s.
Experts abroad are particularly focused on capacity improvements. While HBM is fast, its capacity is limited and comes with a high price tag. In large language model inference, model weights and key-value caches continue to grow. As the number of users increases, the amount of data that must be stored and retrieved within the same GPU also rises.
IT publication Tom's Hardware has evaluated HBF as a new memory layer for AI inference, highlighting its advantage of placing a large memory pool close to processors, which is economically challenging with HBM alone. However, it noted that HBF may struggle to match HBM in terms of latency.
This limitation is seen as a defining factor for HBF's applications. HBM is responsible for the fastest-access data, while HBF serves as a support for read-intensive data that requires large capacity. This is why it is expected to be used first for inference rather than training.
EE Times also views HBF not as a replacement for HBM but as a complementary solution. While it may be disadvantaged in writing speed and durability compared to HBM, it is well-suited for read-centric inference tasks. By placing AI models in HBF and using HBM as a high-speed cache, it is possible to lower memory barriers in a different way.
SK hynix also stands to gain business synergies. The company is a leader in HBM while also having a NAND business. HBF represents a product that can integrate both areas into a single AI infrastructure solution, meaning it can expand the memory footprint within customer systems compared to selling HBM alone.
Kim Cheon-seong, head of SK hynix, stated at FMS 2026, "As AI utilization rapidly spreads, there is a need to redesign the overall data processing structure." This is why HBF is regarded as a technology that redefines the boundaries between memory and storage.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.