University of California, Berkeley and FuriosaAI researchers published a study characterizing high-bandwidth flash for large-language-model serving.

Semiconductor Engineering reports that researchers from the University of California, Berkeley and FuriosaAI published “Characterizing High Bandwidth Flash for LLM Serving.” The subject is high-bandwidth flash (HBF) for large-language-model serving, rather than an announced commercial hardware deployment.

The supplied abstract explains that serving requires substantial memory for model weights and key-value (KV) caches. It identifies memory capacity and bandwidth as increasingly important bottlenecks as models grow and context lengths increase.

The source is a truncated abstract and does not include measured throughput, latency, energy use, cost, or comparisons with high-bandwidth memory. It therefore supports the study’s publication and stated research problem, but not a quantified claim that HBF outperforms or replaces a particular memory technology.