SK hynix and Sandisk to Standardize Next-Gen HBF Memory
SK hynix and Sandisk have announced a joint effort to standardize the next generation of high-bandwidth memory, called HBF. The standardization will occur under the Open Compute Project (OCP) umbrella. The new memory is designed to support the demanding workloads of future AI, machine learning, and cloud computing systems.
- HBF is a new memory tier positioned between high-performance HBM and high-capacity SSDs, built on NAND flash technology instead of the DRAM used in HBM. This design targets the growing AI inference market, where both processing speed and the ability to handle large datasets efficiently are critical. - The first generation of HBF aims for a read bandwidth of 1.6 TB/s and a capacity of 512GB per 16-die stack, which is 8-16 times the capacity of current HBM solutions at a comparable cost. In simulations running a Llama 3.1 405B parameter model, HBF's performance was within 2.2% of a theoretical HBM with unlimited capacity. - The collaboration leverages SK hynix's market leadership in HBM, where it holds over 50% market share, and Sandisk's expertise in NAND flash technology. Key technologies from Sandisk include its BiCS 3D NAND and CBA (CMOS directly Bonded to Array) wafer bonding process. - Sandisk has established a technical advisory board to guide HBF's development, which includes industry leaders like UC Berkeley Professor David Patterson, a distinguished engineer at Google, and Raja Koduri, founder of Oxmiq Labs and former Intel executive. - Sandisk is targeting the second half of 2026 for the first samples of HBF memory, with devices incorporating the new memory expected to be available for sampling in early 2027. Future generations are planned to exceed 3.2 TB/s in bandwidth with stack capacities up to 1.5TB. - The standardization effort is organized as a dedicated workstream within the Open Compute Project (OCP), an open-source hardware initiative founded by Meta in 2011 to drive scalable and efficient data center designs. This approach aims to accelerate broad adoption and build a supportive ecosystem. - HBF is designed to address what is known as the "memory wall," the increasing gap between the rapid performance scaling of AI processors and the comparatively slower growth in memory bandwidth from traditional DRAM architectures. - Unlike DRAM-based HBM, HBF is non-volatile, allowing it to retain data without power and eliminating the need for constant power refresh cycles, which contributes to its power efficiency.
