On August 23, 2023, the 7th Artificial Intelligence Conference 2023, co-hosted by Elecfans and elexcon Shenzhen International Electronics Fair, was held in Shenzhen. Themed "Forge Ahead with AI, Embrace Diverse Intelligence", the conference brought together top industry experts, scholars and enterprise representatives to explore new development opportunities for the artificial intelligence industry.

Tang Qi, Director of AI Inference Engine at Axera, was invited to the conference and delivered a keynote speech entitled Deployment of Transformer Visual Large Models on the Edge Side. He shared the practical implementation of deploying Transformer visual models based on Axera Smart Compute mixed-precision NPU, and officially released the developer kit — Axera Pi Pro. The 4th Artificial Intelligence Excellence and Innovation Awards were presented at the same time. With its strong independent R&D capabilities and outstanding implementation achievements in the edge AI track, Axera's AX650N was awarded the "Most Innovative Value Product Award".

Rapid Popularization of End-side Intelligence, Transformer Visual Model Applications Reach a New Height

Since 2015, when the object recognition accuracy of AI models surpassed that of humans for the first time in the image classification competition on ImageNet, the artificial intelligence industry has officially entered a new era of navigation. In recent years, with the advancement of chip manufacturing processes and the trend of AI model miniaturization and lightweighting, the local deployment of AI models in products such as robotic vacuums, home cameras and smart speakers has become a reality, and end-side intelligence has been popularized.

"So far, the cloud-edge-end complementary computing power framework based on traditional CNN models has been completed," stated Tang Qi in his speech. In the process of AI models migrating from the cloud to the edge and end sides, the Transformer network structure has been continuously evolving and has attracted widespread attention with the boom of ChatGPT. Reviewing the development history of Transformer visual models, from the application of semantic segmentation models based on Transformer network structure in autonomous driving scenarios to the all-in-one segmentation model SAM and the foundational visual large model DINOv2, the applications of Transformer visual models are reaching new heights.

Committed to building world-leading artificial intelligence vision chips, Axera focuses on the R&D of high-performance, low-power AI processor chips for the edge and end sides, and deploys its business in three core markets: smart cities, intelligent driving and AIoT. Driven by the upgrading and transformation of smart cities, the large-scale implementation of L2/L2+ intelligent driving applications, and the growing demand for intelligence of terminal devices, Axera will continue to invest in edge perception intelligent chips and build foundational capabilities in perception and computing.

Accelerate AI Development and Implementation Efficiency, Connect the Digital and Physical Worlds through Vision

Axera's ability to cover all three application scenarios — smart cities, intelligent driving and AIoT — relies on the company's two self-developed core technologies: Axera Smart Vision AI-ISP and Axera Smart Compute mixed-precision NPU.

Axera Smart Vision AI-ISP combines deep learning algorithms with traditional ISP processing units, acting as the "eyes" of universal intelligence to perceive more information, especially meeting the rigid demand for low-light full color. It features six technical highlights: AI Starlight Full Color, AI HDR Imaging, AI Multi-spectral Fusion, AI Anti-shake, AI Scene Enhancement and AI Multi-sensor Fusion. It can comprehensively improve image quality and achieve superior visual effects in various driving scenarios such as dark roads, tunnel exits and bumpy sections.

As Axera's another core self-developed technology, Axera Smart Compute mixed-precision NPU supports three computing precisions — INT4/INT8/INT16 — and boasts prominent advantages of high performance, low cost and ease of use. Notably, drawing on market feedback from the first two generations of products and combining the development trend of cutting-edge industry technologies, the third-generation Axera Smart Compute mixed-precision NPU has further enhanced the types of supported operators and mixed computing with different precisions, optimized the internal memory scheduling mechanism, and can efficiently support both CNN and Transformer network models. Pulsar2, the toolchain developed for Axera Smart Compute 3.0 NPU, has further improved the functionality and ease of use of the toolchain. It supports model deployment on mainstream deep learning training frameworks, and two different quantization operations — PTQ and QAT — to meet the quantization and tuning requirements of various scenarios.

Deploying Transformer models on Axera's NPU is extremely simple and efficient. Taking the SwinT model as an example: directly download the ONNX version of SwinT from PyTorch's official ModelZoo without any modifications to the model or operators, and then realize quantization compression and compilation of SwinT, achieving an ultra-high energy efficiency ratio of 199 FPS/W. AX650N, which was awarded the "Most Innovative Value Product Award" at the conference, is a high-performance chip product launched by Axera this year. Based on the third-generation NPU, it features high performance, high accuracy, easy deployment and low power consumption for Transformer deployment on the edge and end sides, making it an unrivaled Transformer implementation platform in the industry.

Build a Chip Developer Ecosystem, Axera Pi Pro Empowers the Implementation of Visual Large Models

At the conference, Axera officially launched its developer kit — "Axera Pi Pro". Co-created with hardware ecosystem partners, this developer kit enables community developers to experience the convenient deployment of visual large models on the edge and end sides at a low cost. Equipped with AX650N, Axera Pi Pro features high computing power and ultra-strong codec capabilities, which can meet the industry's demand for high-performance edge intelligent computing, realize applications such as video structuring, behavior analysis and status detection, and efficiently support both CNN models and Transformer visual large models.

Along with the launch of Axera Pi Pro, Axera also released the community version of the new-generation AI toolchain Pulsar2, which integrates three functions: model quantization, model compilation and model deployment, as well as abundant development documents. These resources facilitate users to quickly conduct product prototype verification and secondary development, and help users create greater value in fields such as smart cities, intelligent transportation, smart education and intelligent manufacturing.

Up to now, Axera has mass-produced four generations of multiple vision perception chips for different industrial fields, and adheres to the vertical integration of algorithms, chips and products to provide partners with full-stack solutions and help customers achieve the rapid implementation of the latest technologies. Facing the rapid development of the artificial intelligence industry, Axera will continue to pursue technological innovation to improve chip performance and reduce costs, making edge and end-side intelligence more accessible to all, and ultimately fulfilling its mission of "Democratizing AI for a Better Life".