SanDisk and SK hynix team up on high bandwidth flash, offering up to 16x more memory for AI GPUs than HBM

SanDisk and SK Hynix join forces, planning 16x memory boost with new HBF tech tailored for AI GPUs.

: SanDisk has partnered with SK Hynix to pioneer High Bandwidth Flash (HBF) technology aimed at revolutionizing the NAND Flash industry. This new memory innovation is expected to offer up to 16x the capacity of traditional High-Bandwidth Memory (HBM) at similar price points, greatly benefiting AI GPU applications. SanDisk's HBF combines the capabilities of 3D NAND and HBM, targeting AI tasks that demand high bandwidth and large memory capacities. Positioned to enhance both data centers and edge computing, this collaboration anticipates commercial HBF products by early 2027.

SanDisk and SK Hynix have announced a collaborative effort to develop a new form of memory technology known as High Bandwidth Flash (HBF). This innovation seeks to significantly expand the storage capabilities of NAND Flash memory by utilizing a hybrid design that integrates the large storage capacities of 3D NAND chips with the fast data transfer rates associated with High-Bandwidth Memory (HBM). This collaboration intends to create standardized specifications for HBF, which is essential for creating a supportive ecosystem for this new technology.

The partnership aims to address the growing demands of the AI sector, where high-performance computing systems require both fast and expansive memory solutions. As explained by Alper Ilkbahar, SanDisk's Executive Vice President and Chief Technology Officer, HBF technology is engineered to meet the scalability needs arising within the AI industry. The collaboration with SK Hynix, a leading semiconductor manufacturer, is expected to expedite the implementation of HBF solutions across the market.

HBF technology is designed to vastly outperform current memory solutions by enabling up to 16 times more storage capacity than existing HBM at a comparable cost. By vertically stacking eight layers of HBF, a single chip could potentially reach a capacity of 4 terabytes, making it particularly valuable for applications requiring substantial memory, such as AI model training and inference. This advancement suggests that AI models like GPT-4 might eventually be stored directly within a GPU’s VRAM, eliminating the need for external storage devices such as SSDs or network resources.

Hyun Ahn, president of SK Hynix, expressed optimism regarding the potential of HBF technology to advance next-generation technologies. The Korean company is focused on forming a standardized, high-volume, high-performance memory solution. SK Hynix plans to actively contribute to the development and standardization of HBF, with an aim to bring innovative storage solutions to market sooner.

To oversee the progress and market strategies of HBF, a Technical Advisory Board will be formed, consisting of key stakeholders from within the collaborative effort. This board will be responsible for steering the direction of HBF development and deploying effective market strategies. Samples of HBF chips are expected to be available by the second half of 2026, with commercial products likely launching in early 2027, targeting AI inferencing equipped environments, ranging from data centers down to smaller enterprise settings.

Sources: SanDisk, SK Hynix, TechSpot

TechnologyCloudinnovation