Artificial intelligence has fueled demand for faster and larger memory systems, but today’s hardware still forces designers to balance speed, capacity, and cost. Sandisk and SK hynix believe High Bandwidth Flash could help bridge that gap, and the companies have now taken an important step toward making the technology a reality.
The two companies announced the release of the first High Bandwidth Flash technical specification through the Open Compute Project. Rather than introducing a shipping product, the specification establishes a common framework that hardware vendors can use when designing future AI systems around HBF.
Google and Tenstorrent joined the consortium during the standardization effort, lending additional industry support beyond the two memory manufacturers.
Unlike traditional SSDs, HBF is designed to sit much closer to AI processors, occupying a new layer between expensive High Bandwidth Memory and conventional flash storage. The goal is to provide substantially more capacity than HBM while delivering much higher bandwidth than an SSD, giving AI inference systems another option for storing the enormous datasets required by large language models.
The published specification outlines the host interface, electrical requirements, reliability guidance, packaging recommendations, and software behavior needed to support HBF devices. SK hynix also disclosed additional technical details, including support for capacities up to 512GB using 8-high and 16-high NAND stacks, along with bandwidth tiers ranging from approximately 0.4TB/s to 3.0TB/s.
The specification adopts the industry-standard UCIe chiplet interface, allowing HBF to connect with different processor types, including CPUs and GPUs.
Because the specification is being released through OCP, companies across the industry can build compatible products instead of relying on proprietary implementations. That could help accelerate adoption if HBF proves useful in real-world AI infrastructure.
Still, this announcement is primarily about standardization rather than performance. Neither company announced commercial HBF products, published independent benchmarks, nor demonstrated production systems using the technology. Whether HBF ultimately becomes a common component in AI servers will depend on how hardware vendors embrace the standard over the next several years.
Alongside the HBF announcement, SK hynix previewed its upcoming tenth-generation 375-layer V10 4D NAND technology at the Future of Memory and Storage conference. The company says the new NAND improves power efficiency by 2.5 times compared to its previous generation and plans to begin mass production of enterprise SSDs based on the technology early next year.
If HBF gains traction, it could reshape how AI servers organize memory by adding a new tier between today’s ultra-fast HBM and traditional flash storage. That promise is still unproven, but publishing an open specification is a meaningful first step toward building an ecosystem around the technology.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz