Storage Reliability of Multi-bit Flash Oriented to Deep Neural Network

The storage reliability of multi-bit flash is of vital importance for the flash based deep neural network (DNN). In this work, the critical concerns correlated with the storage reliability (I
_d
distribution and data retention) of multi-bit flash and its impacts on the DNN are investigated for the first time. The key achievements include: (1) A dynamic drain-voltage (V
_d
) programming method is proposed to achieve adequately tight (error rate <; 0.5%) and spaced drain-current (I
_d
) distributions in a reasonable time, which leads to comparable accuracy with the software (only 0.25% loss) for the CIFAR-10 recognition. (2) The statistical I
_d
evolution of 16 states over time at different temperatures is studied in a 1Mb flash array, and a physical model is developed to characterize the retention of multi-bit flash. (3) Leveraging the physical model, the device and system co-design is proposed to enhance the reliability of the flash based DNN significantly.