SMART Paths for Latency Reduction in ReRAM Processing-In-Memory Architecture for CNN Inference
share
This research work proposes a design of an analog ReRAM-based PIM (processing-in-memory) architecture for fast and efficient CNN (convolutional neural network) inference. For the overall architecture, we use the basic hardware hierarchy such as node, tile, core, and subarray. On the top of that, we design intra-layer pipelining, inter-layer pipelining, and batch pipelining to exploit parallelism in the architecture and increase overall throughput for the inference of an input image stream. We also optimize the performance of the NoC (network-on-chip) routers by decreasing hop counts using SMART (single-cycle multi-hop asynchronous repeated traversal) flow control. Finally, we experiment with weight replications for different CNN layers in VGG (A-E) for large-scale data set ImageNet. In our simulation, we achieve 40.4027 TOPS (tera-operations per second) for the best-case performance, which corresponds to over 1029 FPS (frames per second). We also achieve 3.5914 TOPS/W (tera-operaions per second per watt) for the best-case energy efficiency. In addition, the architecture with aggressive pipelining and weight replications can achieve 14X speedup compared to the baseline architecture with basic pipelining, and SMART flow control achieves 1.08X speedup in this architecture compared to the baseline. Last but not least, we also evaluate the performance of SMART flow control using synthetic traffic.
READ FULL TEXT