High Bandwidth Flash Faces Narrow Use Cases, OXMIQ Says at Hot Chips
OXMIQ Labs presented at Hot Chips 2026 that High Bandwidth Flash (HBF) cannot replace High Bandwidth Memory (HBM) for the vast majority of workloads. The technology may serve as a specialized memory tier for large, relatively cold datasets, but for other applications it could be counterproductive. The HBF specification includes three performance grades, yet its practical applicability remains highly restricted.
OXMIQ's Hot Chips presentation modeled a 72-GPU rack running a 1-trillion-parameter model to illustrate HBF's trade-offs. An HBF-only configuration boosted capacity to 294.9 TB versus 20.7 TB for HBM-only, but aggregate bandwidth fell from 1,584 TB/s to 922 TB/s. A hybrid approach offered 89.3 TB with bandwidth ranging from 279 TB/s to 1,418 TB/s depending on workload.
The HBF specification defines three performance grades, with Grade 3 reaching 3.072 TB/s via a 32 GT/s UCIe 2.0 interface while retaining 512 GB capacity. Because HBF relies on 3D NAND, it supports 4KB writes and page sizes, limiting its usefulness to large, relatively cold datasets where capacity matters more than speed.
If HBF finds adoption only in narrow niches, AI infrastructure costs may remain tied to expensive HBM, potentially slowing the rollout of very large model deployments for smaller organizations. However, for workloads involving massive but infrequently accessed datasets, HBF could reduce memory costs meaningfully, enabling more efficient use of accelerator clusters. The technology's limited applicability may also influence how chip designers allocate resources, potentially delaying broader memory innovations if expectations outpace practical use cases.