The rapid development of large models has undoubtedly made automated interaction between the physical and digital worlds possible, representing a profound AI-driven transformation of social production and daily life. As these technologies mature, computing costs have also decreased significantly. Today, when enterprises adopt AI, their primary concern is achieving higher-quality service responses with cost-effective computing power, which demands greater economic efficiency and environmental sustainability from computing resources.

In his presentation, Liu Jianwei provided an in-depth analysis of AI technology’s evolution from concept to productivity revolution, and explained how Edge AI Inference power has become the core driver for democratizing embodied intelligence. He explained that AI represents an entirely new programming paradigm, with programming approaches shifting from rule-driven to data-driven methods, and computation primarily based on high-dimensional tensors. In this context, compared to the energy losses caused by data movement in GPU inference, NPUs can more precisely direct the energy consumption generated by computing toward actual data computation, making them more cost-effective and better aligned with AI development’s economic requirements. This principle also represents the original intention behind Axera’s continuous iteration of its AXNeutron NPU.

The widespread adoption of AI infrastructure is fundamentally driven by the need for cost efficiency. Taking computing economic efficiency as an example, Liu Jianwei explained that users must assess “how much AI performance they can get for each yuan spent” (measured by metrics like image processing frame rates and token generation throughput). Axera’s specialized AI processors, built on the AXNeutron NPU technology, achieve this through an “operator instruction set + dataflow microarchitecture” design. This approach delivers energy efficiency 10 times higher than traditional GPUs, bringing AI computing costs down to levels that industries can actually afford.

When discussing embodied intelligence development scenarios, Liu noted that Embodied AI and assisted driving share common technological roots, both following Edge AI Inference’s “perception + decision-making + execution” framework. This shared foundation allows technology costs to be better distributed across applications, enabling Embodied AI to be deployed more rapidly and cost-effectively in commercial scenarios spanning manufacturing, services, and other industries.

Bringing down the price of intelligent infrastructure has become a core objective for all computing chip companies. To achieve a high-performance, cost-effective intelligent computing system, it is essential to construct an operator-based instruction set architecture, which forms the foundation for the underlying microarchitecture design space. At the forum, Liu Jianwei presented Axera’s latest chip for edge intelligence, AXNeutron AX8850. The chip integrates an eight-core A55 CPU, a 24 TOPs@INT8 high-performance NPU, and supports H.264/265 encoding and decoding, 64-bit LPDDR4x, eMMC v5.1/SPI Flash, dual Ethernet ports, one USB3.0 port, two USB2.0 ports, and dual HDMI outputs, enabling users to realize greater value in Edge AI Inference for consumer electronics intelligent upgrades, and other applications.

According to Liu Jianwei, AX8850 offers two core advantages: first, it features a complete operator instruction set that can adapt to evolving AI algorithm requirements; second, it uses a high-performance heterogeneous multi-core architecture with programmable dataflow microarchitecture. The high-efficiency tensor cores are optimized for tensor computation and work seamlessly with vector cores to comprehensively optimize Transformer and BEV algorithms while supporting flexible programming. Combined with SDMA and heterogeneous multi-core hardware coordination, plus rich CV operators for preprocessing and postprocessing, this architecture delivers comprehensive perception performance improvements. The integrated heterogeneous multi-core design optimizes data dependencies through hardware scheduling, minimizing energy waste. On the toolchain side, the company’s software stack, refined through years of mass production experience, enables rapid model deployment, eliminating the need for developers to start from scratch and significantly reducing development barriers.

In specific use cases, AX8850’s environmental perception capabilities can execute semantic segmentation and continuous video analysis in real-time. Meanwhile, multimodal interaction capabilities based on image-to-text technology significantly enhance the naturalness and intelligence of human-machine interaction. Additionally, multi-sensor fusion technology compensates for pure visual perception limitations through cost-effective solutions, making applications in high-precision scenarios such as service robots and assisted driving more targeted and effective.

Flexible scenario adaptation and strong privacy protection make Edge AI Inference better suited to support Embodied AI deployment needs. The AXNeutron NPU series delivers high energy efficiency, strong usability, and rich ecosystems for Edge AI Inference applications, allowing edge devices like robots to better perceive their environment and interact more efficiently with their surroundings. This reflects Axera’s ongoing commitment to its corporate mission “AI for All, AI for a Better World.”