SanDisk and SK hynix Launch HBF Memory Standard to Compete with HBM

iconCryptoBriefing
Share
AI summary iconSummary
SanDisk and SK hynix unveiled High Bandwidth Flash (HBF), a new memory standard combining NAND flash with advanced wafer bonding. The first open specs were released August 3-4, 2026, via the Open Compute Project. HBF targets AI-inference devices, offering 512GB per 16-die stack and up to 3TB/s bandwidth. Samples arrive late 2026, with deployment planned for early 2027. Altcoins to watch may react as fear and greed index shifts with hardware advancements.

SanDisk, working alongside SK hynix, has introduced a new memory standard called High Bandwidth Flash (HBF) that combines NAND flash with advanced wafer bonding techniques to deliver read bandwidths that rival High Bandwidth Memory (HBM) while packing in dramatically more storage capacity.

The first open technical specifications for HBF were released on August 3-4, 2026, through the Open Compute Project. That’s just six months after the standardization effort kicked off in February 2026.

What HBF actually does

HBF is designed to sit between traditional SSDs and HBM in the memory hierarchy. SSDs offer plenty of capacity but relatively slow bandwidth. HBM delivers blistering speed but in small packages.

Advertisement

On a Llama 3.1 405B model benchmark, HBF’s read bandwidth performance came within 2.2% of unlimited-capacity HBM. The technology offers 8 to 16 times more capacity than HBM at comparable cost and power consumption.

The spec supports up to 512GB per 16-die stack and bandwidth tiers reaching up to 3TB/s. It’s also designed to be compatible with HBM4-like footprints and power profiles.

The road to market

SanDisk and SK hynix signed a memorandum of understanding in August 2025, giving both companies roughly six months to align on technical direction before the formal standardization process began in February 2026.

Google and Tenstorrent are among the major tech players participating in the effort. SanDisk expects to have sample HBF products available in the second half of 2026, with integration into AI-inference devices targeted for early 2027.

Why the memory wall matters now

Large language models like Meta’s Llama 3.1 405B, with hundreds of billions of parameters, need to keep enormous volumes of data accessible at high speed during inference. HBM handles the speed requirement well but becomes prohibitively expensive when you need terabytes of it.

The technical specifications released through the Open Compute Project include detailed guidelines for die stacks, read and write operations, and electrical interfaces, which other vendors can use to build compatible products.

SK hynix is the dominant supplier of HBM, so its participation in a technology that could partially substitute for HBM suggests a strategic hedge. SanDisk, meanwhile, gets to move up the value chain from commodity flash storage into a higher-margin, performance-critical product category.

Disclaimer: The information on this page may have been obtained from third parties and does not necessarily reflect the views or opinions of KuCoin. This content is provided for general informational purposes only, without any representation or warranty of any kind, nor shall it be construed as financial or investment advice. KuCoin shall not be liable for any errors or omissions, or for any outcomes resulting from the use of this information. Investments in digital assets can be risky. Please carefully evaluate the risks of a product and your risk tolerance based on your own financial circumstances. For more information, please refer to our Terms of Use and Risk Disclosure.