The 22nd China International Semiconductor Expo (IC China 2025) was held at the China National Convention Center from November 23-25, 2025. As part of the event, Liu Jianwei, Co-founder and Vice President of Axera, was invited to deliver a keynote speech at the Forum for Artificial Intelligence and Large Model Chips. His presentation, titled “AI Processors Drive Edge and Terminal Intelligence,” shared the company’s practices and insights in the terminal and edge AI chip sector.
The main content of his speech is as follows:
As AI technology matures, its applications are expanding from the cloud to the terminal, gradually taking on more sophisticated roles in perception and control. From ssmart vehicle to intelligent factories and smart homes, AI is enabling cross-scenario intelligent connectivity through collaborative applications. Founded in 2019, Axera has consistently focused on developing terminal and edge AI chips to promote AI accessibility and accelerate intelligent infrastructure at the edge and terminal.
AI not only delivers more intelligent experiences but also represents “a stable and reliable intelligent foundation like water and electricity.” Particularly with the rise of large model technology, AI has enabled large-scale replication of human experience and knowledge, transforming processes that previously depended on manual judgment into standardized industrial products. This significantly improves production efficiency and drives intelligent upgrades across various industries, representing a cost-driven productivity revolution.
Facing such widespread AI applications, how should we build the supporting computing infrastructure? Axera offered its answer by examining the evolution of computing architecture. Just as GPUs emerged from CPUs to meet graphics computing demands, current AI applications’ requirements for computing density have made general-purpose processors suboptimal in terms of energy efficiency and cost-effectiveness. Therefore, designing dedicated processing architectures for AI has become essential for driving large-scale development of terminal and edge intelligence.
Axera’s solution is a dedicated AI processor architecture featuring the combination of operator instruction set and dataflow DSA microarchitecture. This architecture treats AI programs as computational graph structures, directly implementing common operators as processor execution instructions, thereby providing more efficient hardware-level support for high-dimensional tensor data flow and computation. Compared to traditional GPU architectures, this design significantly reduces power consumption and latency from data movement, concentrating more resources on productive computation. This notably improves energy efficiency and cost-effectiveness while being more environmentally sustainable.
AI processor evaluation must account for both explicit and implicit costs: chip price and power consumption constitute explicit costs, while development efficiency, toolchain maturity, and ecosystem completeness affect users’ R&D investment and time-to-market. Axera not only focuses on improving chip performance but also commits to building comprehensive software toolchains and development ecosystems, providing complete resources from model conversion and integrated development to reference designs, and leveraging developer communities and open-source platforms (such as Hugging Face, GitHub, etc.) to lower development barriers.
In terms of hardware form, Axera has collaborated with ecosystem partners to launch various products, including development boards, computing cards, AI boxes, and edge servers, addressing requirements across different stages from R&D to deployment. Meanwhile, the company has adapted various mainstream algorithms through partnerships with algorithm ecosystem providers. This diversified approach enables customers across different industries to select solutions flexibly based on their specific application needs, achieving rapid deployment and implementation.
On-site demonstrations showcased a series of terminal AI applications powered by Axera’s AI chips, including image-to-text, text-to-image, video understanding, and multimodal tasks. These capabilities can be widely applied to high-frequency scenarios, such as behavior recognition, fall detection, and multimedia content generation and summarization, fully demonstrating the core advantages of terminal AI in real-time processing, privacy protection, and cost-effectiveness.
Axera will continue to pursue its mission of “AI for All, AI for a Better World,” collaborating with partners through proprietary AI chips featuring “high energy efficiency, low power consumption, and superior cost-effectiveness,” along with open, win-win ecosystem solutions. The company aims to accelerate intelligent infrastructure at the edge and terminal, continuously empowering various industries with AI, and making intelligent technology truly accessible to every household.
