Sandisk proposes HBF for AI at the edge

Sandisk Corp. is making significant strides in the realm of memory technology with its innovative approach to 3D-NAND flash memory. The company is positioning this technology as a potential replacement for DRAM-based HBM in AI inference applications. This move marks a strategic shift for Sandisk, which was spun off from Western Digital in February 2025 with a focus on delivering flash memory products and exploring disruptive memory technologies.

During the Sandisk Investor Day just before the spin off, Alper Ilkbahar, the incoming senior vice president of memory technology, introduced the concept of high bandwidth flash and 3D Matrix Memory. He highlighted the development of high bandwidth flash (HBF), which optimizes NAND flash for bandwidth rather than die area and cost. This innovative architecture involves dividing the NAND memory array into multiple mini arrays that can be accessed in parallel, leveraging Kioxia BICS 3D-NAND technology to stack these mini arrays vertically.

According to Ilkbahar, the use of HBF has resulted in the creation of 16-layer R&D memories with significantly higher capacities compared to traditional HBM, all while maintaining a similar price point. He emphasized that major AI players are providing input into the development of HBF, which has the potential to revolutionize datacenter GPUs and expand their applications to AI-capable smartphones and other edge devices.

One of the key advantages of HBF highlighted by Ilkbahar is its ability to address the growing memory requirements of advanced AI models. For instance, a contemporary LLM like GPT4 with 1.8 trillion parameters and 16-bit weights demands 3.6Tbytes of memory. By utilizing HBF, a single GPU could potentially accommodate the entire model, eliminating the need for extensive data movement and enhancing overall efficiency.

Looking ahead, Ilkbahar envisions HBF playing a crucial role in enabling AI on smartphones by overcoming memory, performance, and power constraints. By leveraging HBF, future AI models with billions of parameters and specific weight requirements could be housed within a single die, paving the way for enhanced AI capabilities at the edge. While acknowledging that HBF is not a direct replacement for HBM, Sandisk is actively working on establishing an open-standard interface based on existing electrical interfaces with minimal protocol changes.