" 国产算力多芯片、多架构并存的当下，谁为碎片化买单？ " 作者丨赵之齐 编辑丨包永刚 2025年12月12-13日，第八届GAIR全球人工智能与机器人大会在深圳·博林天瑞喜来登酒店正式启幕。 作为AI产学研投界的标杆盛会，GAIR自2016年创办以来，始终坚守"传承+创新"内核，始终致力于连接 技术前沿与产业实践。 在人工智能逐步成为国家竞争核心变量的当下，算力正以前所未有的速度重塑技术路径与产业结构。13日 举办的"AI算力新十年"专场聚焦智能体系的底层核心——算力，从架构演进、生态构建到产业化落地展开 系统讨论，试图为未来十年的中国AI产业，厘清关键变量与发展方向。 在大会上， 清程极智联合创始人、产品副总裁师天麾，带来了题为《智能算力的适配、优化和服务》的主 题演讲。 在国产算力从"能用"走向"好用"的关键阶段，AI Infra正从幕后走到台前。 师天麾给出的判断颇为直接：国产算力利用率的瓶颈，更多在于软件生态与系统级优化能力。 无论是围绕国产芯片的全栈推理引擎自研、通过纯软件方式提前跑通FP4等低精度路线，还是在MaaS （模型即服务）市场中用评测、路由与统一接口"消除信息差" ，师天麾试图回答的， ...

Core Insights - The ChinaSC 2025 conference focused on the theme of "Intelligent Computing Power, Large Models, New Economy," discussing the technological trends and policy directions in China's computing power development [1] - The event featured the release of the "2025 China High-Performance Computing Performance TOP100 Ranking" and the "2025 China Computing Power Leading Enterprises Award" [2] - The AIPerf500 international AI computing power ranking was updated, highlighting the advancements in AI training and inference performance [3][4] Industry Developments - The conference emphasized the importance of AI as a driving force for transformation across various industries, with efficient AI computing power being crucial for the development and implementation of large models [5][6] - The establishment of the Ankang Intelligent Computing Center aims to become a key hub for computing power in Western China, with a target of building a 20,000P cluster [7] - The integration of AI and HPC (High-Performance Computing) was discussed, with innovations in software and algorithms being essential for overcoming structural bottlenecks in traditional HPC applications [8] Technological Innovations - The AIPerf ranking introduced new metrics for evaluating AI computing systems, focusing on training capabilities and inference performance [3][4] - Companies like Beijing Super Cloud Computing Center and Alibaba Cloud were recognized for their high-performance AI computing systems [3] - The development of liquid cooling technology was highlighted as a key innovation for enhancing computing power across various applications [9][10] Strategic Collaborations - A strategic cooperation agreement was signed between the Ankang High-tech Zone Management Committee and the China Intelligent Computing Industry Alliance to foster collaboration in infrastructure, ecosystem development, and technology transfer [11] - The conference also recognized outstanding contributions in the field, awarding several individuals and companies for their achievements in computing power technology [12][13] Future Outlook - The China Intelligent Computing Industry Alliance plans to continue its efforts in promoting the development of the computing power industry, focusing on practical applications and addressing technological challenges [14] - The conference concluded with a strong emphasis on the need for collaboration and innovation to drive the growth of the computing power economy in China [15]

Core Insights - The company, Pingao Co., Ltd. (688227.SH), has disclosed its technological advancements in the AI sector, focusing on software and algorithm optimization to reduce reliance on high-performance hardware, particularly in the context of overseas high-end chip bans [1] - The trend in the industry indicates that AI software capabilities are improving, leading to lower hardware performance requirements, as modern AI algorithms show increased tolerance for hardware errors and noise [2] - Pingao's approach combines domestic chips with software optimization, allowing less powerful domestic chips to achieve higher performance through innovative software solutions [3] Industry Trends - The industry consensus is shifting towards software optimization to enhance the efficiency of domestic chips, with various domestic enterprises and research institutions investing in this direction [2] - Notable advancements include Tsinghua University's "Bagua Furnace" training system and "Chitu" inference engine, which optimize the efficiency of domestic computing power [2] Company Solutions - Pingao's "Pingyuan AI All-in-One Machine," developed in collaboration with Jiangyuan Technology, exemplifies the company's strategy, achieving a 30% increase in response speed for the DeepSeek-R1 model and a 2.5 times improvement in energy efficiency compared to mainstream GPUs [3] - The company has developed the BingoAIInfra intelligent computing power scheduling platform, which enhances the utilization of domestic hardware by allowing precise management of GPU resources [4] Ecosystem Layout - Pingao is building a comprehensive "hardware-software-ecosystem" system to ensure the sustainable development of its technology, including strategic investments in domestic chip companies and collaboration on optimizing inference algorithms [5] - The company’s Pingao Cloud operating system supports a wide range of domestic heterogeneous chip servers and applications, creating a self-controlled ecosystem that mitigates risks associated with overseas technology limitations [5] Conclusion - In the context of rapid digital economy and AI industry growth, Pingao's innovative approach not only achieves technological breakthroughs but also provides a viable path for mainstream AI applications to transition from reliance on overseas high-end hardware to domestic chips, thereby driving the autonomous development of domestic AI computing power [6]

Core Viewpoint - The rapid evolution of large models has shifted the focus from the models themselves to systemic issues such as slow inference, unstable training, and data migration challenges, which are critical for the scalable implementation of technology [1] Group 1: Key Issues in Domestic AI - Domestic AI faces challenges including computing power adaptation, system fault tolerance, and data compliance, which are essential for its practical application [1] - The AICon conference will address seven key topics focusing on the infrastructure of domestic AI, including native adaptation of domestic chips for inference and cloud-native evolution of AI data foundations [1] Group 2: Presentations Overview - The "Chitu Inference Engine" by Qingcheng Jizhi aims to efficiently deploy FP8 precision models on domestic chips, overcoming reliance on NVIDIA's Hopper architecture [4] - Huawei's "DeepSeek" architecture will discuss performance optimization strategies for running large models on domestic computing platforms [5][6] - JD Retail's presentation will cover the technical challenges and optimization practices for high throughput and low latency in large language models used in retail applications [7] - Alibaba's session will explore the design and future development of reinforcement learning systems, emphasizing the complexity of algorithms and system requirements [8] - The "SGLang Inference Engine" will present an efficient open-source deployment solution that integrates advanced technologies to reduce inference costs [9] - Ant Group will share insights on stability practices in large model training, focusing on distributed training fault tolerance and performance analysis tools [10] - Zilliz will discuss the evolution of data infrastructure for AI, including vector data migration tools and cloud-native data platforms [11]