University of Oxford researchers reported that High-Bandwidth Flash (HBF) provides 16-fold more capacity per stack than High-Bandwidth Memory at comparable bandwidth.

University of Oxford researchers published a technical paper titled “Hardware-Managed Heterogeneous High-Bandwidth Memory and Flash in LLM Inference Systems.” The work examines a hybrid memory architecture for large-language-model inference.

The paper describes High-Bandwidth Flash, or HBF, as a denser alternative to High-Bandwidth Memory, or HBM. It reports 16 times more capacity per stack at comparable bandwidth.

HBM is high-speed stacked memory used close to processors and AI accelerators, while HBF in this research refers to stacked flash designed to provide high bandwidth with substantially greater capacity. The comparison concerns capacity per stack and bandwidth, not processor compute performance.

The paper investigates hardware management of heterogeneous HBM and HBF resources for inference systems. The supplied abstract excerpt says replacing HBM with HBF can address capacity constraints, but the source data does not provide full latency, energy, cost or end-to-end inference benchmark results.