A Hardware Automated Domain-Specific Flash Memory System for Emerging Applications
Shine Kim, Jae Wook Lee · 2022
NAND flash-based SSDs have become a key component in various emerging applications because of their high performance, large capacity, and cost-efficiency. A recent work proposes a flash memory system (FMS) to enable data-parallel training of deep neural networks (DNNs). Although the FMS effectively replaces the high-cost HBM DRAM of the conventional DNN accelerator with low-cost NAND flash memory using fully hardware automated storage data path, its benefits as a NAND-based SSD and the detailed effects of the automated data path are yet to be evaluated. In this paper, we evaluate the performance of FMS as a storage device. In addition, unlike the previous work, we apply the automated storage data path to various multi-level cell (MLC) NAND-based FMSs to demonstrate the effectiveness of the automated data path for future emerging applications requiring both higher capacity and performance. The performance evaluation results show that the automated data path of FMS improves storage performance by up to 5x compared to the conventional SSD. Finally, FMS, freed from an SSD firmware bottleneck via the automated data path, proves its performance and capacity are scalable as the number of NAND chips increases.