NewsBite analysis · AI, Hardware
wccftech
HBF Is For AI Model Weights, HBM Is For KV Cache
SanDisk, in collaboration with SK hynix, is developing a new memory standard called High-Bandwidth Flash (HBF) designed to alleviate the bottlenecks faced by High Bandwidth Memory (HBM) in AI models. This initiative will enable the storage of model weights in HBF while reserving HBM for Key-Value (KV) caches, which are crucial for enhancing processing speeds. With up to 512GB storage and bandwidth reaching between 0.4TB/s and 3TB/s, HBF offers a cost-effective alternative to traditional HBM, potentially leading to extended model context without compromising accuracy.
MonitorConfidence: HighMid term horizon
Published 5 Aug 2026, 00:20 UTC4 min read
Visit original article ↗Lead image — wccftech
Summary
TL;DR
- SanDisk developing High-Bandwidth Flash (HBF) with SK hynix.
- HBF aims to alleviate memory bottlenecks in AI processing.
- Allows for separating model weights storage from KV caches.
- HBF offers 512GB storage and bandwidth of 0.4TB/s to 3TB/s.
- Potential to enhance long context models while preserving accuracy.
Interesting facts
- HBF leverages parallelism to enhance NAND read speeds significantly.
- HBF may be cheaper than HBM, unlocking economies of scale.
- Model weights can now shift to HBF while KV cache remains on HBM.