SK hynix Inc. and Sandisk Corporation launched a global standardization initiative for a next generation memory solution called HBF, or high bandwidth flash, at a consortium kick off event held at Sandisk headquarters in Milpitas, California on the 25th. The companies stated that by making HBF an industry standard they intend to lay the foundation for the entire Artificial Intelligence ecosystem to grow together, and they plan to start formal standardization work through a dedicated workstream under the Open Compute Project.
The move responds to a shift in the Artificial Intelligence industry from training, which focuses on creating large language models, to inference, which accelerates actual Artificial Intelligence services to users. As the number of users of Artificial Intelligence services increases rapidly, fast and efficient memory becomes critical for handling data intensive inference workloads. Existing memory structures cannot meet high capacity data processing and power efficiency requirements simultaneously at the inference stage, and HBF technology is presented as a way to address these limitations.
HBF technology is defined as a new memory layer positioned between ultra fast high bandwidth memory and high capacity solid state drives. The design aims to fill the gap between the performance of high bandwidth memory and the capacity of solid state drives, while ensuring both capacity expansion and power efficiency required for Artificial Intelligence inferencing. In the proposed architecture, high bandwidth memory handles the highest level of bandwidth and HBF serves as a supporting layer that complements it, creating a more balanced hierarchy between compute, memory bandwidth, and storage capacity for Artificial Intelligence workloads.
