Shulou(Shulou.com)11/24 Report--

On August 23, 2023, the Seventh artificial Intelligence Conference 2023, jointly sponsored by Electronic Fever Friends and elexcon Shenzhen International Electronics Show, was held in Shenzhen. With the theme of "ambitious and capable", this conference brings together top experts, scholars and enterprise representatives to look for new opportunities for the development of artificial intelligence industry.

Tang Qi, director of Ai Xin Yuanzhi AI reasoning engine, was invited to attend the conference, delivered a keynote speech on "deployment of Transformer visual model on the edge side", shared the landing practice of deploying Transformer visual model based on Ai Xintong yuan mixed precision NPU, and officially released Ai Xin Pai Pro, a developer kit. The fourth artificial Intelligence Excellence Innovation Award was awarded at the same time. With its strong self-research strength and outstanding landing results in the marginal AI track, Ai Xin Yuanzhi AX650N was awarded the "most innovative value Product Award".

Rapid popularization of end-to-side intelligence Transformer visual model application to a new height

Since the object recognition accuracy of AI model surpassed that of human for the first time in ImageNet image classification competition in 2015, the artificial intelligence industry has officially entered a new navigation era. In recent years, with the progress of the manufacturing process of the chip industry and the trend of miniaturization and lightweight of AI model, AI model has become a reality in the local deployment of floor sweeping robots, household cameras, intelligent speakers and other products, and end-to-side intelligence has been popularized.

"up to now, the main line of computing power of cloud, edge, and end triangle complementary based on the traditional CNN model has been completed," Tang Qi said in his speech. In the process of AI model migration from cloud to end-to-side, Transformer network structure is also constantly improving, and has attracted more and more attention with the popularity of ChatGPT. Reviewing the development history of Transformer vision model, from the application of semantic segmentation model based on Transformer network structure to autopilot scene, to the application of SAM and pedestal vision model DINOv2,Transformer vision model is moving to a new height.

Committed to creating the world's leading artificial intelligence vision chip, Aixin Yuanzhi focuses on high-performance, low-power edge-side and end-to-side artificial intelligence processor chip development, layout smart city, intelligent driving and AIoT three major business markets. Based on the upgrading of Smart City, the large-scale landing of L2 / L2 + intelligent driving applications, and the expansion of intelligent terminal equipment, Aixin Yuanzhi will continue to invest in edge-aware smart chips and continue to build the basic capabilities of perception and computing.

Speed up the landing efficiency of AI development and visually connect the digital world and the physical world

From smart city to intelligent driving to AIoT, the reason why Ai Xin Yuanzhi can cover all three major application scenarios depends on the company's self-developed Ai Xin Zhiye AI-ISP and Ai Xin Tong Yuan mixed precision NPU.

AI-ISP combines deep learning algorithm with traditional ISP processing unit to perceive more information, especially the rigid requirement of dark light and full color, as the "eyes" of everything intelligence. It has six technical highlights: AI starlight full color, AI HDR imaging, AI multi-spectral fusion, AI anti-shake, AI scene enhancement, AI multi-sensor fusion, which can comprehensively improve the image quality and achieve better visual effects in dark road, tunnel exit, bumpy road and other traffic scenes.

NPU, as another core self-developed technology, supports three kinds of calculation accuracy of INT4 / INT8 / INT16, and has the outstanding advantages of high performance, low cost and easy to use. It is worth mentioning that, drawing on the market feedback of the first two generations of products and combined with the development trend of cutting-edge technology in the industry, the third generation Aixintong mixed precision NPU further enhances the operator to support mixed computing with different kinds and different precision, optimizes the internal memory scheduling mechanism, and supports CNN network and Transformer network model with high efficiency. On the other hand, Pulsar2, which is developed for Aixin Tongyuan 3.0NPU, further improves the function and ease of use of the tool chain, supports the model deployment of the mainstream deep learning training framework, supports two different quantitative operations of PTQ and QAT, and satisfies the quantitative tuning function of various scenarios.

The deployment of the Transformer model on Ai Xin Yuanzhi NPU is very simple and efficient. Take the SwinT model as an example: download SwinT's ONNX model directly from the official ModelZoo of PyTorch without any model or operator modification to achieve the quantitative compression and compilation of SwinT, achieving a high energy efficiency of 199 FPS / W. At the meeting, AX650N, which was awarded the "most innovative value Product Award", is a high-performance chip product launched by Aixin Yuanzhi this year. Based on the third-generation NPU, Transformer deployed on the edge side and end-side has the characteristics of high performance, high precision, easy deployment and low power consumption, and is the industry's leading Transformer landing platform.

Create the chip developer Eco-Love Core send Pro to help the big visual model land

At the conference, Aixin Yuanzhi officially launched the developer suite-"Ai Xin Pai Pro", a developer suite jointly created by hardware ecological partners, which is intended to provide community developers with a low-cost experience of easy deployment of large visual models on the edge and end-to-side. Aixin Pro is equipped with AX650N, with high computing power and strong codec ability, which can meet the needs of the industry for high-performance edge intelligent computing, realize video structure, behavior analysis, state detection and other applications, and efficiently support CNN model and Transformer visual model.

While Ai Xin sends Pro for sale, Ai Xin Yuanzhi will also release the community version of Pulsar2, a new generation of AI tool chain with three-in-one functions of model quantification, model compilation and model deployment, as well as rich development documents, so as to facilitate users to quickly carry out product prototype verification and secondary development, and help users to play greater value in smart city, intelligent transportation, intelligent education, intelligent manufacturing and other fields.

At present, Aixin Yuanzhi has mass-produced more than four generations of visual perception chips used in different industries, and adheres to the vertical integration of algorithms, chips and products to provide partners with full-stack solutions. to help customers achieve the rapid landing of the latest technology. Facing the rapid development of artificial intelligence industry, Aixin Yuanzhi will also improve chip performance, reduce costs, make edge intelligence more affordable, and finally realize the mission of "inclusive AI to create a better life" through continuous technological innovation.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.