The development of on-device edge computing technology is heating up as related makers are seeking technology breakthroughs to address the concerns of privacy and protection of sensitive date when applying deep learning under cloud applications and to lower costs for enterprises to adjust its internal IT infrastructure to cope with cloud environments; or to challenge the restrictions caused by latency, Internet bandwidth limits and insufficient network infrastructure for development of smart applications such as computer vision (CV) and self-driving automobiles.
In consideration of power consumption and performance, deep learning attached to end devices currently focuses on inference, which is normally done by embedding a training model in the form of chip into terminal-end devices. However, Google has been implementing its Federated Learning as a decentralized data training architecture, so that terminal devices will be no longer working mainly on deep learning inference.
Federated Learning also allows end devices to perform small-scale data training reducing bandwidth demand and the burden of cloud servers, while protecting privacy and sensitive data. It also helps improve the efficiency and speed of the circulation from training to inference, and therefore enhancing edge computing.
While Qualcomm has rolled out its neural process engine (NPE) to enable AI on smartphones via heterogeneous computing, the incorporation of AI chips into system on chip (SoC) products has become a prevailing trend. The introduction of integrated AI chips on application processors by Apple and Huawei will encourage other chipmakers including Qualcomm and MediaTek to follow suit, Digitimes Research believes.
Meanwhile, a number of IP providers including Synopsys, CEVA, Cadence, Verisilicon all have come out with DSP (digital signal processor)-based embedded solutions, which some of them are made of a 16nm process, for the CV market, where they will meet competition from Intel and Nvidia, especially after Intel has acquired Mobileye and Movidius to cross into autonomous driving and image recognition sectors.
