SK hynix and SanDisk Introduce High Bandwidth Flash to Bridge the AI Memory Gap
SK hynix and SanDisk are working to solve one of the biggest bottlenecks in modern AI hardware: the massive performance gap between High Bandwidth Memory and traditional enterprise SSD storage. Their answer is a new memory standard called High Bandwidth Flash, or HBF, designed to sit between ultra-fast HBM and high-capacity SSDs.
As artificial intelligence models continue to grow, especially large language models with hundreds of billions of parameters, AI servers need more than just raw computing power. They require memory systems that can move huge amounts of data quickly while still offering enough capacity to store increasingly large AI workloads. HBM delivers extreme bandwidth, but it is expensive and limited in capacity. SSDs offer much larger storage, but they cannot match the speed needed for many AI inference tasks.
High Bandwidth Flash aims to fill that gap.
Unlike HBM, which uses stacked DRAM, HBF uses stacked NAND flash memory. This allows it to deliver much higher capacity while still offering significantly better bandwidth than standard SSD-based storage. The goal is not to replace HBM entirely, but to complement it. HBM can continue handling the most immediate compute-heavy workloads, while HBF can store larger AI models and datasets closer to the processor for faster access.
According to the specifications introduced by SK hynix and SanDisk, HBF can support configurations of up to 512GB by stacking NAND dies in eight-layer or 16-layer designs. Bandwidth can range from 0.4TB/s to 3TB/s, placing it in a much faster class than conventional storage while still offering the capacity benefits of NAND flash.
This could become especially important for AI inference. For example, a large language model with 500 billion parameters may require around 500GB of memory if each gigabyte holds roughly one billion parameters. Storing that entire model in HBM would be extremely expensive and impractical for many companies. With HBF, AI systems could keep large models in a high-capacity, high-bandwidth memory tier, reducing the need to constantly move data from slower storage.
Another key advantage of High Bandwidth Flash is support for UCIe, or Universal Chiplet Interconnect Express. This allows HBF to connect more flexibly with CPUs, GPUs, and other chiplet-based architectures. As AI hardware moves toward more modular designs, this compatibility could help manufacturers build more efficient systems that combine compute, memory, and storage more closely than before.
SK hynix has described HBF as part of a broader effort to blur the line between memory and storage. The company believes this approach can help create new system architectures that improve overall efficiency, particularly in AI servers where data movement is one of the most expensive and time-consuming challenges.
The timing is important. AI infrastructure is under increasing pressure as companies deploy larger models and demand faster inference performance. GPUs are often limited not only by compute capability but also by how quickly they can access the data they need. A faster intermediate memory layer like HBF could help reduce GPU offloading delays and improve real-world AI performance.
Meanwhile, Samsung is also pushing forward in the high-speed storage space for AI servers. The company has been developing advanced NAND-based solutions and working on high-layer NAND flash technologies, including designs that may reach up to 500 stacked layers through future V10 and V11 NAND production. These efforts are aimed at meeting the growing need for ultra-fast data access in AI infrastructure.
The competition now appears to be forming around different approaches to the same problem: how to give AI systems faster access to larger amounts of data without relying entirely on costly HBM. SK hynix and SanDisk are betting on High Bandwidth Flash as a new standard, while other major memory players are developing their own high-performance NAND and memory expansion technologies.
If HBF gains industry support, it could become an important part of next-generation AI accelerators, data center GPUs, and inference servers. Its ability to combine high capacity, strong bandwidth, and chiplet-friendly connectivity may make it attractive for companies building large-scale AI systems.
For now, the biggest question is adoption. A new memory standard needs support from chipmakers, server manufacturers, and cloud infrastructure providers before it can become widely used. But with AI models growing rapidly and memory bottlenecks becoming more severe, the demand for a solution like High Bandwidth Flash is clear.
The coming months should reveal whether HBF can gain momentum and become a key technology in the future of AI computing.
