Sophon BM1880 is SoC ASIC chip for Deep Learning inference acceleration focusing on edge application. BM1880 TPU can provide 1TOPs performance for 8-bit integer operation and up to 2TOPs by enabling Winograd convolution acceleration. The special designed TPU schedule engine is very powerful, and can supply extremely high bandwidth data flow efficiently to all the Tensor Processor cores. 2MB on chip memory enables best programing flexibility for performance optimization and data reuse.

A powerful deep-learning model compiler and software SDK are provided to end users. The mainstream deep learning frameworks (such as Caffe / Tensor flow) could be ported to BM1880 platform easily. All the popular CNN / RNN / DNN network topologies can be greatly accelerated on BM1880 platform.

Sophon BM1880 is the multi-functional platform for AI. It can be defined as the co-processor to the host for Deep Learning Inference acceleration. It can also be used as the host to receive video, images or other data from the Ethernet or USB interface, then perform inference and other computer vision tasks. Host SOC can send the raw image files of each video frame or the motion jpeg files to BM1880, BM 1880 can then inference the images and return the results back to the host.