On November 22nd, at the 2019 annual meeting of China's IC design industry, CEVA technical support manager Zhang Wei gave a speech on “Leading from product to product-AI technology to lead the development of camera equipment”.
Zhang Wei said that CEVA is a pioneer in the field of AI and computer vision, and also a leader, achieving four firsts. In 2011, the company launched the image and vision processor, and equipped it with a rich software library. In 2015, it released the computer vision DSP CEVA-XM6 and CDNN software framework; in 2018, the first generation AI processor architecture was launched; in 2019 Second generation AI processor architecture NeuPro-S.
Today's AI computer vision applications are everywhere, including autonomous vehicles, smart phones, surveillance cameras, consumer cameras, AR/VR, robotics and industrial applications.
According to Zhang Wei, NeuPro-S is one of the good choices for AI computer vision. NeuPro-S splits, detects, and classifies neural networks in video and imagery on edge devices, significantly improving system-aware performance. Compared to CEVA's first-generation AI processors, NeuPro-S improves performance by an average of 50%. Bandwidth and power consumption are reduced by 40% and 30%, respectively. The NeuPro-S series includes the NPS1000, NPS2000, and NPS4000, which are pre-configured processors with 1000, 2000, and 4000 8-bit MACs per cycle to meet the needs of different application scenarios.
It is worth noting that in addition to AI real-time processing, the fully programmable CEVA-XM6 vision DSP integrated in the NeuPro-S architecture can simultaneously process images, computer vision and general DSP workloads. Zhang Wei stressed that this allows customers and algorithm developers to take advantage of CEVA's rich library of image and visual software, including the CEVA-SLAM software development kit for 3D mapping, CEVA-CV and CEVA-VX software for computer vision development. Library, wide-angle image software suite (including de-warping, video stitching and image data sensor fusion technology).
In conjunction with the release of NeuPro-S, CEVA also launched the industry's first deep neural network compiler technology CDNN-Invite API, which optimizes neural network inference firmware to support heterogeneous collaborative processing of NeuPro-S kernel and custom neural network engine through a unified interface. .
Zhang Wei pointed out that NeuPro-S, together with the CDNN-Invite API, is an ideal choice for any vision-based device that requires edge artificial intelligence processing.