SK Hynix and SanDisk have begun standardizing High‑Bandwidth Flash, an intermediate memory between HBM and SSD

SK Hynix and SanDisk announced the launch of a new memory tier – High Bandwidth Flash (HBF) – as part of the Open Compute Project

What is HBF
* Positioning:

HBF is a “mid‑range” level between ultra‑fast HBM memory and high‑capacity SSD.

It fills the gap between the high throughput of HBM and the large capacity of SSD, providing both volume expansion and energy efficiency needed for AI workloads.

* Role in architecture:

Unlike HBM, which provides core high speed, HBF acts as an auxiliary memory layer that complements the overall system.

How SanDisk sees HBF
Characteristic | SanDisk’s View | Technology
---|---|---
Memory type | NAND memory using BiCS NAND and CBA (CMOS Direct Bonded to Array) technology. |
Capacity | 8–16× larger than HBM at comparable throughput. |
Cost | Roughly comparable to HBM in typical system configurations.

First‑generation stack
* Read throughput: up to 1.6 TB/s (256 Gb/s per die).
* Total stack capacity (16 dies): up to 512 GB.
* Physical parameters: the stack is designed to fit optimally with size, power consumption and height of HBM4 – the integration model is similar to modern HBM stacks on AI accelerators.

Performance
Internal tests showed that HBF differs by only 2.2 % from “unlimited‑capacity HBM” under a workload based on an 8‑bit Llama 3.1 405B configuration.

*Important:* the comparison uses a hypothetical HBM without capacity limits, so it does not directly measure the benefit of larger capacity.

Data retention
HBF retains data without power – unlike DRAM, which requires constant power to refresh data.

Future plans (Gen 2 and Gen 3)
Generation | Target read throughput | Stack capacity | Power consumption (vs. Gen 1)
---|---|---|---
Gen 2 | > 2 TB/s | up to 1 TB | 0.8× of Gen 1
Gen 3 | > 3.2 TB/s | up to 1.5 TB | 0.64× of Gen 1

What’s next
* Standardization: SK Hynix and SanDisk have already identified HBF as the next step within the Open Compute Project.
* Deployment: A precise schedule has not yet been announced.

Thus, HBF promises to become a new “mid‑range” memory layer that combines the speed of HBM with the volume of SSD while remaining energy efficient and compatible with existing AI accelerators.