SK Hynix and Sandisk Unveil High Bandwidth Flash to Rival HBM for AI

SK Hynix and Sandisk have unveiled a new High Bandwidth Flash (HBF) initiative under the Open Compute Project. The move positions HBF as a new AI memory tier between HBM and SSD storage.

The announcement follows SK hynix’s recent HBM3e supply deal with Microsoft, further highlighting the race to secure advanced memory for AI accelerators.

A New Memory Tier for AI Inference

According to VideoCardz, SK hynix and Sandisk are leading a dedicated workstream within the Open Compute Project (OCP) to standardize High Bandwidth Flash.

HBF targets AI inference workloads and aims to deliver far higher capacity than High Bandwidth Memory (HBM) without falling back to slower, traditional storage-like access patterns.

Sandisk describes HBF as a NAND-based memory solution tailored specifically for AI accelerator systems. The concept focuses on balancing bandwidth, capacity, and power efficiency.

Performance and Capacity Targets

According to the companies, HBF could provide roughly 8× to 16× the capacity of HBM in target AI designs.

First-generation specifications include:

• Up to 1.6 TB/s read bandwidth
• 256Gb per die
• 16-die stacks reaching up to 512GB capacity

HBF stacks are designed to match HBM4 in footprint, stack height, and power profile. This alignment should simplify integration into next-generation AI accelerators.

Internal simulations suggest HBF can achieve performance within 2.2% of “unlimited-capacity HBM” in inference workloads. The test reportedly used an 8-bit Llama 3.1 405B model.

The comparison assumes unlimited HBM capacity, meaning it primarily evaluates bandwidth behavior rather than the real-world capacity advantage HBF offers.

Built on Advanced NAND and 3D Stacking

HBF relies on Sandisk’s BiCS NAND roadmap and its CMOS directly Bonded to Array (CBA) architecture.

The companies are developing proprietary 3D stacking techniques aimed at 16-die configurations. These designs target improved thermals and reduced warpage, both critical in dense AI hardware environments.

Unlike DRAM-based memory such as HBM, HBF is non-volatile. It does not require refresh power, which may translate into better energy efficiency in large AI clusters.

Roadmap: Scaling Beyond 3 TB/s

The roadmap outlines significant performance growth:

• Gen2 aims for over 2 TB/s read bandwidth
• Gen3 targets up to 3.2 TB/s read bandwidth

Future stack capacities could scale to 1TB and even 1.5TB per stack. The companies also project lower power consumption, with Gen2 and Gen3 potentially dropping to 0.8× and 0.64× of Gen1 power levels.

No official deployment timeline has been announced. For now, both firms treat OCP standardization as the next major milestone.

The announcement lands as the AI memory race accelerates. Samsung recently revealed HBM4, while NVIDIA’s upcoming Rubin architecture is widely expected to drive further HBM4 demand.

If HBF gains traction, it could reshape how AI accelerators handle inference workloads, offering a middle ground between ultra-fast HBM and high-capacity SSD storage.