3 6 Ke·2026-01-29 02:28Core Viewpoint - Alibaba Group has launched its high-end AI training and inference chip, Zhenwu 810E, marking a significant step in its AI strategy with the unveiling of its full-stack AI architecture "Tongyun Ge" [1][3]. Group 1: Product Launch and Features - The Zhenwu PPU has been deployed in multiple large-scale clusters on Alibaba Cloud, serving over 400 clients including State Grid, Chinese Academy of Sciences, Xiaopeng Motors, and Sina Weibo [3]. - The Zhenwu 810E chip features a self-developed parallel computing architecture and ICN inter-chip interconnection technology, with a single card supporting 96GB HBM2e memory and an interconnection bandwidth of 700GB/s [4]. - The chip is designed for AI training, inference, and autonomous driving, and has been successfully deployed in multiple large-scale clusters on Alibaba Cloud [4]. Group 2: Performance Comparison - The Zhenwu PPU's overall performance surpasses NVIDIA's A800 and mainstream domestic GPUs, and is comparable to NVIDIA's H20, with an upgraded version reportedly outperforming NVIDIA's A100 [6]. Group 3: Industry Reception and Strategy - Industry professionals have praised the Zhenwu PPU for its excellent stability and cost-effectiveness, contributing to a positive reputation in the market [7]. - Alibaba aims to develop "Tongyun Ge" into an AI supercomputer, leveraging its self-developed chips, leading cloud services, and renowned open-source models to achieve maximum efficiency in training and deploying large models [7]. Group 4: Advantages and Applications - The Zhenwu 810E boasts six key advantages: fully self-developed, high performance, strong interconnectivity, high usability, large-scale validation, and integration of chip and cloud [7][8]. - The chip's software stack supports various applications, including autonomous driving, AI training, AI inference, and multi-modal models, demonstrating strong performance across different scenarios [9]. Group 5: Company Background and Vision - Alibaba Cloud was established in 2009, and the chip company Pingtouge was founded in 2018, with a strategic focus on AI model research starting in 2019, culminating in the complete layout of the "Tongyun Ge" full-stack AI [10]. - Pingtouge aims to innovate chip infrastructure for the AI era, enhancing data center capabilities and efficiency through integrated hardware and software solutions [10].