Core Insights - The core viewpoint of the news is that Zhonghao Xinying's "Shanai®" TPU architecture AI chip has successfully adapted to Baidu's open-source multimodal mixture of experts model ERNIE-4.5-VL-28B-A3B, demonstrating the efficiency of domestic TPU architecture in supporting cutting-edge models and establishing a new ecosystem paradigm of "domestic innovative chip architecture + domestic open-source large models" [1][2]. Company Overview - Zhonghao Xinying was established in 2018 by Yang Gongyifan, a core developer of Google's TPU chip, along with a team of AI hardware and software design experts from major tech companies like Google, Microsoft, and Samsung. The company has a comprehensive methodology for chip design and optimization across various process technologies from 28nm to 7nm, with over 70% of its workforce dedicated to R&D [1]. Product Performance - The "Shanai®" TPU architecture AI chip, after nearly five years of development, features fully controllable IP cores, self-developed instruction sets, and computing platforms. It surpasses renowned overseas GPU products by nearly 1.5 times in AI large model computing scenarios while reducing energy consumption by 30%. The chip employs Chiplet technology and 2.5D packaging to achieve performance leaps under the same process technology, supporting interconnection of 1024 chips for linear scaling in large model computations [1]. Model Adaptation - The ERNIE-4.5-VL model, which has a total parameter count of 28 billion and an active parameter count of 3 billion, utilizes a heterogeneous mixture of experts (MoE) architecture. It excels in cross-modal understanding and generation, as well as long text processing, making it suitable for various applications such as intelligent navigation and visual customer service [2]. Technical Integration - The integration of Zhonghao Xinying's "Shanai®" TPU AI chip with the ERNIE-4.5-VL model showcases enhanced parallel processing capabilities, improving computation speed and accuracy for complex tasks. The chip's reconfigurable multi-level storage and near-memory computing design effectively support the model's performance in handling multimodal data [3]. Application and Development - The technology team at Zhonghao Xinying has successfully executed multiple complex multimodal tasks using the "Shanai®" TPU AI chip, demonstrating its capability to provide stable and powerful computational support for large models. The chip meets the demands of both large-scale model training and real-time inference tasks, further optimized through close collaboration with Baidu's PaddlePaddle framework [4]. Future Directions - Yang Gongyifan, the founder and CEO of Zhonghao Xinying, stated that the successful adaptation validates the feasibility of collaborative innovation between domestic computing power and models. The company plans to deepen its technical collaboration with Baidu to implement hardware acceleration solutions for a full range of models from 3 billion to 424 billion parameters, aiming to provide more efficient and reliable domestic AI infrastructure [4].