作者｜郑松毅 编辑｜杨锦 PPU是专用加速处理器，像计算界的 "专业工坊"。对比CPU/GPU，PPU舍弃了通用性而专攻单一任务（如 AI 大模型 训推、游戏物理模拟），在特定目标任务上更高效、更节能。 搜狐科技从官网获悉，"真武"PPU采用自研并行计算架构和片间互联技术，配合全栈自研软件栈，实现软硬件全自 研。其内存为96G HBM2e，片间互联带宽达到700 GB/s，可应用于AI训练、AI推理和自动驾驶。目前，阿里巴巴已 将"真武"PPU大规模用于千问大模型的训练和推理。 出品｜搜狐科技 1月29日上午，平头哥官网悄然上线一款名为"真武810E"的AI芯片，此前被央视《新闻联播》曝光的阿里自研芯片PPU 正式亮相。这是通义实验室、阿里云和平头哥组成的阿里巴巴AI黄金三角"通云哥"首次浮出水面。 阿里巴巴正在将"通云哥"打造成一台AI超级计算机，它同时拥有全栈自研芯片平头哥、安全稳定的阿里云，以及性能 优异的开源模型"千问"，可以在芯片架构、云平台架构和模型架构上协同创新，从而实现在阿里云上训练和调用大模 型时达到最高效率。 在2025年云栖大会，阿里云智能集团首席技术官周靖人向搜狐科技介绍，"阿里和谷歌 ...

Core Insights - Huawei unveiled the complete development roadmap for its Ascend AI chips at the Global Connectivity Conference, establishing a self-controlled computing foundation for China's AI industry amid manufacturing constraints [2] - The domestic AI chip ecosystem is evolving with multiple players like Alibaba, Baidu, Tencent, and ByteDance adopting different strategies for market entry, creating a multi-path domestic substitution landscape [3][6] Huawei's Asymmetric Innovation Path - Huawei acknowledges that its single-chip computing power still lags behind Nvidia but opts for an asymmetric innovation path, focusing on system architecture and interconnection technology to compensate for performance gaps [3] - The key breakthrough is the Lingqu interconnection protocol, which addresses long-distance, high-reliability, and low-latency challenges, achieving significant improvements in reliability and bandwidth [3] Future Roadmap for Ascend Chips - Huawei's roadmap includes the launch of Ascend 950PR in Q1 2026, focusing on inference pre-fill scenarios, and Ascend 950DT in Q4 2026, which will shift to SIMD/SIMT architecture for general AI tasks [4] - The Ascend 960, set for Q4 2027, will double specifications, while the Ascend 970 in Q4 2028 will achieve significant increases in computing power and memory access bandwidth [4] Competitive Edge of Huawei's SuperNodes - Huawei's Atlas 950 SuperPoD and Atlas 960 SuperPoD support 8192 and 15488 Ascend cards respectively, leading in key metrics compared to Nvidia's upcoming NVL144S [4] - The Atlas 950 SuperCluster and Atlas 960 SuperCluster boast computing power exceeding 500,000 and 1 million cards respectively, marking a significant advancement in the industry [4] Software and Ecosystem Development - Huawei is enhancing its software and ecosystem by open-sourcing the CANN compiler and virtual instruction set interface, with plans to open-source the Mind series application enablement suite by the end of 2025 [5] - The company aims to build a competitive technology alliance against Nvidia's CUDA ecosystem through the open release of Lingqu 2.0 technology specifications [5] Other Players in the AI Chip Ecosystem - Alibaba is diversifying its AI chip strategy with self-developed inference chips and ARM architecture CPUs, while Baidu's Kunlun series has gained market recognition with significant contracts [8] - Tencent is optimizing software-hardware collaboration through cloud services and investments in AI chip companies, while ByteDance is engaging in supply chain partnerships to procure domestic AI chips [8] Policy and Patent Support - The Chinese government has implemented policies to support the chip industry, aiming for a 70% self-sufficiency rate by 2025, with various tax incentives for chip manufacturers [10] - Patent accumulation is crucial for technological breakthroughs, with Huawei holding over 150,000 global patents, including significant numbers in AI and communication fields [11][12] Conclusion - The advancements in Huawei's Ascend chips and the collaborative efforts of other domestic players are driving the growth of a self-controlled AI chip ecosystem in China, supported by robust government policies and patent strategies [12]