SK hynix and Sandisk Propose New 'HBF' Memory Standard
SK hynix and Sandisk have formed a joint workstream under the Open Compute Project (OCP) to begin global standardization of a next-generation memory technology called 'HBF'. The new memory layer is designed to fit between HBM and SSDs to improve scalability and power efficiency for AI inference infrastructure.
- The HBF standard targets read-intensive AI inference workloads by offering memory capacity up to 16 times larger than HBM at a comparable cost, with a single stack potentially reaching 512GB. - HBF utilizes NAND flash chips stacked vertically with Through-Silicon Vias (TSVs), an architectural approach similar to HBM, to achieve high bandwidth. However, its access latency is in the microsecond range, slower than HBM's nanosecond latency, and it has a lower write endurance of approximately 100,000 cycles. - The proposed implementation is a hybrid architecture where HBF serves as a high-capacity tier for storing large AI models, while HBM acts as a faster cache for frequently accessed "hot data" closer to the GPU. - SK hynix has published simulation results for a hybrid design, termed "H³," using an NVIDIA Blackwell GPU, which showed a performance increase of up to 6.14 times in tokens per second and the ability to handle 18.8 times more simultaneous queries than an HBM-only configuration. - The collaboration is significant as it combines Sandisk's NAND and BiCS technology with SK hynix's deep expertise in HBM manufacturing and GPU interposer integration. - Sandisk, a subsidiary of Western Digital, plans to deliver the first HBF samples in the second half of 2026, with the first AI inference devices incorporating the technology expected in early 2027. - The standardization effort is being conducted as a dedicated workstream within the Open Compute Project (OCP) to foster a multi-supplier ecosystem and accelerate adoption. - Key executives involved in the initiative include Sandisk EVP and CTO Alper Ilkbahar and SK hynix President and Chief Development Officer, Dr. Hyun Ahn.
