Hot Chips 2026: High Bandwidth Flash promises massive bandwidth and capacity, but its usability is extremely limited — new memory format strikes a balance between HBM and NAND flash

At Hot Chips 2026, OXMIQ Labs announced High Bandwidth Flash (HBF) as a high-capacity memory technology. While HBF cannot replace High Bandwidth Memory (HBM) for most workloads, it could serve as a specialized memory tier for large, but relatively inactive datasets. OXMIQ stated that HBF offers eight to sixteen times more capacity than HBM at a comparable cost.

However, the main advantage of HBF lies in its capacity, not in its performance. HBF's performance is a secondary consideration, with the technology supporting read block sizes between 64 bytes and 4 kilobytes, 4KB writes, and 4KB page sizes. When bandwidth becomes the limiting factor, HBF's lower bandwidth becomes a bottleneck, making HBM a more cost-effective choice for maximum inference throughput from a fully utilized rack.

HBF can reduce the number of GPUs required for very large models, such as OXMIQ's Kimi-K3 model with 1.56 TB of weights, of which 93% consists of MoE expert weights. Keeping these experts in HBF and frequently accessed weights in HBM could minimize expert transfers, reduce all-to-all communication, and potentially decrease network traffic.

HBF could also store large KV caches for sparse-attention models, fetching only the necessary data from HBF to HBM during each decoding step.