A Real-Time 1280 × 720 Object Detection Chip With 585 MB/s Memory Traffic

Memory bandwidth has become the real-time bottleneck of current deep learning accelerators (DLA), particularly for high definition (HD) object detection. Under resource constraints, this paper proposes a low memory traffic DLA chip with joint hardware and software optimization. To maximize utilization under bandwidth, we morph fuse detection model into group fusion-ready to reduce intermediate data access. This reduces YOLOv2's feature from 2.9 GB/s 0.15 GB/s. support fusion, our previous based employes unified buffer write-masking simple layer-by-layer processing in fusion group. When compared same PE numbers, implemented TSMC 40nm process supports 1280x720@30FPS consumes 7.9X less external DRAM access energy, 2607 mJ 327.6 mJ.