Fusing In-Storage and Near-Storage Acceleration of Convolutional Neural Networks

Video analytics have a wide range of applications and has attracted much interest over the years. While it can be both computationally energy intensive, video greatly benefit from in/ near memory compute. The practice moving compute closer to continued show improvements performance consumption is seeing increasing adoption. Recent advancements in solid state drives (SSDs) incorporated Field Programmable Gate Arrays (FPGAs) with shared access drive’s storage cells. These FPGAs are capable running operations required by analytic pipelines such as object detection template matching. typically executed using Convolutional Neural Networks (CNNs). A CNN composed multiple individually processed layers which perform various image processing tasks. Due lack resources, layer may partitioned into more manageable sub-layers. sub-layers then sequentially, however some simultaneously. Moreover, cells within FPGA equipped SSD’s being augmented in-storage accelerate workloads exploit intra parallelism layer. To this end we present our work, leverages heterogeneous architectures create an in/near-storage acceleration solution for analytics. We designed NAND flash accelerator, mapped evaluated several benchmarks. how utilize FPGAs, local DRAMs, in-memory SSD workloads. Our work also demonstrates remove unnecessary transfers save latency energy.