
Premium content
Access to this content requires a subscription. You must be a premium user to view this content.

keywords:
approximate nearest neighbor search
3d-nand
neuromorphic computing
This paper demonstrates a highly reliable 3D-NAND-based billion-scale approximate nearest neighbor search (ANNS) accelerator, which can evaluate the Manhatten (L1) distance between the query and multi-billions of instances directly in the SSD. Three approaches (1. “Partial-ON” string-select line (SSL) input, 2. Thermometer code data-encoding scheme , and 3. On-demand SSL re-trimming) are discussed to improve the accuracy and stability. Proposed computational SSD can maintain high recall rate for the SIFT-1B test even after 200℃ baking.
