Axera Announces Mixed Precision NPU “Axera Tongyuan” at ICDIA 2023
On July 13 to 14, 2023, the 3rd China Integrated Circuit Design Innovation Conference and IC Application Expo (ICDIA 2023) took place. At a specialized forum focusing on AIoT and ChatGPT, Axera’s Co-founder and Vice President, Jianwei Liu, delivered a keynote speech titled “Axera AX650N: Empowering Transformer Deployment on Edge and Terminal Devices.” He officially introduced the name for the company’s core mixed-precision NPU technology, known as “Axera Tongyuan,” while showcasing the third-generation SoC, AX650N, and its remarkable capabilities in deploying Transformers on edge and terminal devices.
Axera Unveils Name for Its Mixed Precision NPU, “Axera Tongyuan”
In recent years, the Artificial Intelligence of Things (AIoT) industry has experienced rapid growth, emerging as the premier route for the intelligent upgrade of traditional industries and a key direction for the future development of the Internet of Things (IoT). Since its debut at the end of 2022, ChatGPT and other large models have spurred major tech giants to intensify their efforts, marking the dawn of a new era in artificial intelligence development. With the rapid advancement of AIoT and large AI models, the underlying hardware is faced with ever-increasing demands for data storage, computing power, and graphics processing capabilities.
Axera believes that large models, with their general intelligence capabilities, can lower the costs of AI deployment across various scenarios. In the future, everyone could have a smart assistant on their device. In the development of artificial intelligence, Axera provides the essential infrastructure with its AI chips, delivering critical perception and computational capabilities to enable AI deployment on edge and terminal devices.
On the perception front, Axera is dedicated to enhancing camera clarity, creating a digital gateway to the physical world. On the computing front, Axera is committed to enabling cameras to interpret what they see, providing a robust computational foundation for deploying various AI models on edge and terminal devices. At ICDIA 2023, Axera officially unveiled the name for its mixed precision NPU, “Axera Tongyuan.” This technology aims to provide fundamental computational support for various intelligent algorithms, enabling a deeper understanding of the world on edge and terminal devices, and contributing to a better life.
Specializing in AI perception and foundational edge computing platforms, Axera has been focusing on perception and computing capabilities since its founding in 2019. The company has developed two core technologies: the Axera Zhimou AI-ISP and the Axera Tongyuan NPU. The latter overcomes memory and power consumption barriers, delivering higher effective computing power to support more intelligent algorithms in space- and power-constrained edge and terminal environments, thereby reducing AI deployment costs.
So far, Axera has completed the development and mass production of four generations of chips, gradually deploying them in the smart city, smart driving, and AIoT markets. “These markets all rely on perception and computation as fundamental capabilities, which is why Axera has chosen to focus on them,” said Liu.
High Performance, High Precision, Easy Deployment: Axera AX650N Emerges as the Optimal Platform for Transformer Deployment
When designing and developing AI chips, Axera prioritizes the seamless integration of applications, algorithms, and the NPU. In terms of application, Axera boosts performance through data flow optimization and accelerated front-end/back-end processing. On the algorithm side, Axera enhances hardware utilization with operator acceleration, network microstructure acceleration, and memory optimization. Additionally, the Axera Tongyuan NPU is a heterogeneous multi-core system with a built-in multi-core hardware scheduling mechanism that reduces CPU usage, enabling applications on the system to run faster.
Axera’s AI chips for edge and terminal devices are engineered with a design philosophy that harmonizes applications, algorithms, and NPU optimization, resulting in high performance and low power consumption. Its third-generation SoC, AX650N, exemplifies high computing power and efficiency, making it the optimal platform for deploying Transformers.
Real-world testing shows that Axera’s AX650N excels not only in traditional CNN but also in deploying Transformer networks like SwinT on edge devices. It delivers impressive results with 361 FPS, high accuracy at 80.45%, low power consumption at 199 FPS/W, and easy deployment using the original model with PTQ (Post Training Quantization). Additionally, the AX650N supports low-bit mixed precision. By using INT4, users can significantly reduce memory and bandwidth consumption, effectively lowering the costs for edge and terminal deployment.
Currently, Axera’s AX650N supports a range of Transformer models, including ViT/DeiT, Swin/SwinV2, and DETR. The cutting-edge self-supervised computer vision model, DINOv2, also achieves over 30 FPS on the AX650N. To help developers better deploy Transformers, Axera has introduced the “AxeraPi Pro” development kit based on the AX650N. This kit is designed for the ecosystem community and industry applications, supporting the exploration of more diverse product applications.
Amid the new wave of AI enthusiasm driven by large models, Axera plans to further optimize the AX650N for the Transformer architecture and explore multi-modal Transformer models. Axera aims to accelerate the deployment of edge and terminal intelligence through its core perception and computing technologies, ultimately making AI more accessible to all and improving their quality of life.
