The 3rd China Integrated Circuit Design Innovation Conference & Wuxi IC Application Expo (ICDIA 2023) was held from July 13th to 14th. At the special forum themed "AIoT and ChatGPT", Liu Jianwei, Co-founder and Vice President of Axera, was invited to deliver a keynote speech entitled Axera AX650N: A Transformer Implementation Platform for Edge and End Sides. He officially announced the Chinese name of the company's core technology – the mixed-precision NPU, "Axera Smart Compute", and introduced the leading advantages of the third-generation SoC chip AX650N in deploying Transformer on the edge and end sides.

Axera Announces the Chinese Name of Its Mixed-Precision NPU: Axera Smart Compute
In recent years, the AIoT industry has experienced rapid scale expansion, serving not only as the optimal channel for the intelligent upgrading of various traditional industries but also as a crucial development direction for the Internet of Things in the future. Meanwhile, since its release at the end of 2022, large models represented by ChatGPT have spurred major tech giants at home and abroad to launch relevant deployments, ushering artificial intelligence into a brand-new era of development. It is evident that with the rapid development of AIoT and large AI models, the underlying hardware is facing extremely high requirements in terms of data storage, computing performance and graphics processing capabilities.

In Axera's view, large models possess general cognitive capabilities that can reduce the cost of implementing artificial intelligence in specific scenarios, and in the future, everyone may have an intelligent assistant on their terminal devices. In the course of artificial intelligence development, Axera provides chip infrastructure – that is, delivering foundational perception and computing capabilities through AI chips to facilitate the implementation of AI on the edge and end sides.

At the perception level, Axera is committed to enabling cameras to capture clearer images, thus creating an entrance for the digitalization of the physical world. At the computing level, Axera strives to make cameras "understand" what they capture, and will provide a more powerful computing foundation for the implementation of various AI models on the edge and end sides. On the site of ICDIA 2023, Axera officially announced the official Chinese name of its mixed-precision NPU – "Axera Smart Compute", which is intended to provide basic computing power support for various intelligent algorithms, enable intelligence to gain a deeper understanding of the world on the edge and end sides, and contribute to a better life.

Positioned as a foundational computing power platform for AI perception and edge computing, Axera has focused on building foundational perception and computing capabilities since its establishment in 2019, and independently developed two core technologies: Axera Smart Vision AI-ISP and Axera Smart Compute mixed-precision NPU. The latter breaks through the constraints of the memory wall and power consumption wall, and supports more intelligent algorithms with higher effective computing power under the limited area and power consumption conditions on the edge and end sides, thereby reducing the cost of AI deployment.

Up to now, Axera has completed the R&D and mass production of four generations of multiple chip products, and has gradually rolled out its solutions in three major markets: smart cities, intelligent driving and AIoT. "Perception and computing are regarded as foundational capabilities in all these markets, which is the logic behind Axera's selection of different market segments," said Liu Jianwei.

High Performance, High Accuracy, Easy Deployment: Axera AX650N Becomes the Optimal Transformer Implementation Platform
In the design and development of AI chips, Axera attaches great importance to the in-depth integration of applications, algorithms and NPU. In terms of applications, Axera achieves higher performance through data flow optimization and front-and-back end processing acceleration. In terms of algorithms, Axera improves hardware utilization from an algorithmic perspective through designs such as operator acceleration, network micro-structure acceleration and memory optimization. Meanwhile, Axera Smart Compute mixed-precision NPU is a heterogeneous multi-core system with a built-in multi-core hardware scheduling mechanism that reduces CPU occupancy and enables faster operation of system applications.

Based on the design philosophy of joint optimization of applications, algorithms and NPU, all the edge and end-side AI processor chips launched by Axera feature high performance and low power consumption. Its third-generation high computing power, high energy efficiency ratio SoC chip AX650N has even become the optimal implementation platform for Transformer.

Test data shows that Axera's AX650N not only delivers outstanding performance on traditional CNN networks but also excels in deploying the Transformer network SwinT on the end side, achieving a high performance of 361 FPS, a high accuracy of 80.45%, a low power consumption of 199 FPS/W, and extremely easy deployment with the original model and PTQ quantization. In addition, AX650N supports low-bit mixed precision; if users adopt INT4, it can greatly reduce memory and bandwidth occupancy, thus effectively controlling the cost of deployment on the edge and end sides.

At present, Axera's AX650N has been adapted to Transformer models such as ViT/DeiT, Swin/SwinV2 and DETR, and the most advanced computer vision self-supervised model DINOv2 also achieves a running speed of over 30 frames on AX650N. To help developers make better attempts at the implementation of Transformer, Axera has also launched the "Axera Pi Pro" development kit based on AX650N for the ecological community and industrial applications, facilitating the exploration of more abundant product applications.

Facing the new round of AI upsurge driven by large models, Axera's AX650N will be further optimized for the Transformer structure and explore Transformer models in the multimodal direction. Through its core technologies in perception and computing, Axera aims to accelerate the implementation of intelligence on the edge and end sides, truly realize the democratization of AI, and make people's lives better.