High-bandwidth memory (High-bandwidth memory, HBM) has played a key role in AI accelerators, but capacity limits appear to be driving growing interest in high-bandwidth flash (high-bandwidth flash, HBF).
TechRadar reported on Jan. 23 that HBM is fast but limited in capacity, while HBF offers larger capacity but is relatively slower.
Samsung Electronics and Sandisk plan to integrate HBF into Nvidia, AMD and Google AI products within the next 24 months. SK Hynix is also preparing a prototype, and Kioxia is developing a 5TB HBF module using PCIe Gen 6 x8.
HBF is expected to emerge alongside HBM6 and change memory architecture. HBM6 expands bandwidth and capacity at the same time by connecting multiple HBM stacks through a network, and a "memory factory" concept that processes data directly in HBF is expected to be introduced in HBM7.
In the AI memory market, some assessments say the arrival of HBF represents a fundamental change in how data is processed. There are also forecasts that HBF will surpass the HBM market by 2038.
Chi-gyu Hwang
delight@d-today.co.kr