Computational Storage for 3D NAND Flash Error Recovery Flow Prediction

The Computational Storage paradigm is attracting increasing interest in many applications because of the performance and the energy-efficiency improvement, given by the tight coupling of processing elements with Solid State Drives through proper interconnection fabrics. In this work, we study a computational storage architecture aimed to boost the inference step of an Artificial Neural Network designed to predict the Error Recovery Flow outcome from the 3D NAND Flash memories characterization data. The application has been implemented on the Xilinx Alveo U250 Data center accelerator using a 15 bits fixed point precision, proving a 98.6% prediction accuracy, a performance boost up to 53.5 ×, and two orders of magnitude energy consumption reduction with respect to a CPU-only implementation.