Currently, the development of the electronics industry is facing unprecedented challenges. On the one hand, the complex international environment and intensified market competition have led to intensified involution in the industry; on the other hand, the continuous update and iteration of new technologies have brought brand new opportunities to enterprises. With risks and opportunities coexisting, the upstream and downstream of the industry are still forging ahead with perseverance and continuous innovation, which further highlights the resilience of Chinese semiconductor enterprises.

As a leader in the domestic AI chip field, Axera has deeply laid out end-side applications. Its AI chips, with a network model architecture natively supporting Transformer, fully support the practice of end-side large model deployment. In this regard, Tang Qi, Director of AI Recommendation Engine at Axera, delivered a themed sharing at the conference. Mr. Tang stated that large models are migrating and being deployed from the cloud-side to the end-side. End-side large models have the advantages of strong privacy, convenient offline deployment, and customizability. This has also driven the vigorous development of multimodal large models. However, it should also be noted that current end-side large models still face challenges such as limited options, high costs, high power consumption, and the failure to form a closed-loop of application implementation. To address these issues, Axera has launched a series of solutions, including mass-produced NPUs and easy-to-use software stacks.

Tang Qi revealed that Axera has launched the R&D of the latest Neutron V6p NPU this year. This is an AI chip with computing power of up to 128 TOPS at 8-bit precision per second, which can be adapted to various large models and help realize a variety of AI functions such as text-to-image search, one-click removal, target acquisition, and monocular depth estimation. At the same time, a brand-new development board based on AX630C — Axera Pi 2 — has also been launched, which can facilitate more community developers to deploy large models on the end-side.