SK Hynix & SanDisk Launch HBF AI Memory Standard

🔑 Enhanced Key Takeaways

• •HBF positions as a new memory layer between HBM and SSD, offering HBM-comparable bandwidth with 8-16x greater capacity at similar cost for AI inference.[1][2][4]
• •SanDisk's HBF leverages advanced BiCS NAND technology and proprietary CBA wafer bonding, with a first-generation prototype featuring 16-layer chips up to 512GB per stack.[2][4][5]
• •The collaboration includes a dedicated workstream under the Open Compute Project (OCP) and follows an MOU signed prior to the kick-off event.[1]
• •SanDisk formed a Technical Advisory Board with industry experts, including input from AI players and figures like Raja Koduri, and won 'Best of Show, Most Innovative Technology' at FMS 2025.[2][3][4]

🛠️ Technical Deep Dive

• •HBF uses SanDisk’s BiCS 3D-NAND technology with proprietary CBA (Chip-on-Wafer Bonding) wafer bonding to integrate NAND flash into HBM-like packages.[2][4]
• •First-generation HBF features 16-layer memory chips, enabling up to 512GB capacity per stack while matching HBM bandwidth.[5]
• •Architecture breaks NAND memory arrays into smaller sections operating simultaneously, stacked vertically for higher capacity (8-16x HBM) at comparable cost and power efficiency.[5]
• •SK hynix's related 'LPW NAND' trademark enhances I/O channels while reducing per-channel speed to lower power for AI inference.[5]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline