NetFlex: A 22nm Multi-Chiplet Perception Accelerator in High-Density Fan-Out Wafer-Level Packaging.

NetFlex is a multi-chiplet package (MCP) for CNN-based perception acceleration. With a balanced parallelism for mapping and a flexible scheduling, the NetFlex chiplet supports convolution, deconvolution and fully connected layers of different shapes, sizes and strides at high utilization. NetFlex adopts depth-first stream processing and an efficient streaming interface in a multi-chiplet daisy chain over Advanced Interface Bus. A 22nm NetFlex chiplet was fabricated and measured to achieve 2.14TOPS/W (16b OP) at a nominal voltage of 0.89V and 492.3MHz. A four-chiplet NetFlex MCP was built in a high-density fan-out wafer-level packaging to demonstrate 428FPS for depth estimation and 7723FPS for pose estimation.