High-Bandwidth Flash offers efficient storage for AI model weights
A novel solution to overcome memory bottlenecks in AI inference workloads.
1Memory demand explosion
The proliferation of Large Language Models has created unprecedented memory demands in data centers. As traditional DRAM and HBM solutions face production capacity and cost challenges, the industry is turning to next-gen storage technologies.
2Flash memory evolution
NAND flash memory used in SD cards is being transformed through 3D packaging techniques. High Bandwidth Flash (HBF) aims to break the read speed limits of conventional flash with multi-layer stacked architecture.
3Technical advantages
HBF stands out with energy efficiency and high storage density. Its non-volatile nature prevents data loss during power outages. 16-layer models promise 512GB capacity and 1.6TB/s read speeds.
4Purpose-built for AI inference
Unlike training processes, AI inference workloads use model weights in read-only mode. HBF's slow write speeds aren't a disadvantage here. HBM can be reserved exclusively as high-speed scratchpad memory.
5Industry collaboration
Sandisk and SK Hynix are working on HBF standardization under the Open Compute Project. The initiative launched in 2026 aims to create a new memory tier in data center architectures.
6Future outlook
HBF could democratize AI by reducing infrastructure costs. Expected to enter production in 2027, this technology will play key roles in edge computing and large-scale inference workloads.