Title: 从秀单卡到建体系，WAIC 见证国产算力系统性进化 | BestBlogs.dev
URL Source: https://www.bestblogs.dev/article/4b7f5b15ce?amp%3Butm_medium=feed&%3Butm_campaign=resources&%3Bentry=rss_article_item
Published Time: 2026-07-18 17:08:00
Markdown Content: 81
The article reports on the progress at WAIC 2026 where domestic computing power shifted from single-chip performance competition to full-stack software and hardware system upgrades, including the open-source SAIL software stack, supernode products, and ecosystem collaboration. 
Today 1134 words (about 5 min) View Source →
Sign in to highlight text and take notes as you read. Sign in now
陈杨园 2026-07-18 17:08 上海
2026.07.18
本文字数：1004，阅读时长大约2分钟 **作者 |**第一 财经 陈杨园
“我们始终相信AI时代的真正爆发不是来自某一颗、两颗芯片，而会来自一套更开放、更协调、更高效的全栈策略算力体系。”7月18日，在2026世界人工智能大会（WAIC ）上，平头哥半导体副总裁高慧如此表示。
当天，平头哥开源自研AI软件栈T-Head SAIL（以下简称“SAIL”），这是平头哥真武AI芯片的底层软件，可最大程度释放芯片的算力。
今年以来，智能体的快速应用正在重塑数据中心，带来AI智算的新难题。智能体应用超长上下文多轮推理、多智能体高频协同等场景，带来AI芯片空转等待、算力利用率降低、交互频繁切换导致无效能耗激增等情况，挑战之下，国产算力基础设施也在加速进化。
本届WAIC上，记者关注到，伴随着国产芯片的进展，算力企业已不止于秀单颗芯片，伴随着更多超节点和集群被展出、AI芯片底层软件开源，国产算力正迎来系统级升级。
过去几年，国产算力强化了AI芯片这一“发动机”后，作为“变速箱+传动系统+驾驶系统”的AI软件栈也在加速进化。高慧介绍，最新开源的平头哥SAIL软件架构覆盖OS 层、SDK层与接口层的完整链路，不仅可以释放真武芯片算力，还可高效承接上层应用需求，在软件生态上，SAIL全面兼容主流AI生态，同时适配260多个主流训练与推理框架。
“开源AI软件栈只是开始，平头哥希望与全球开发者、科研机构及产业伙伴，共同打磨工具链、丰富算子生态、优化性能体验。”高慧表示。
值得关注的是，本届WAIC也成为超节点产品集中亮相的舞台。平头哥在大会展出了真武M890×磐久AL128超节点，单柜集成128颗芯片，通过自研ALink互连架构实现Pb/s级带宽与百纳秒级延迟，同时，阿里云灵骏真武 M890 超节点实例在会上发布，这也是阿里云首次通过公共云对外提供超节点形态的 AI 算力服务；华为则展出了超节点Atlas 950 SuperPoD，现场展示1024张昇腾卡，基于超节点的Atlas 950 SuperCluster集群算力规模可扩展至50万卡；此外，还有沐曦股份在大会发布新一代AI超节点产品曦景S600，中兴通讯发布中兴OEX超节点新品。在这背后，面向AI大模型时代，国产算力厂商的核心硬件基座进化明显加速。
国产算力正进入“下半场”竞争。国盛证券在近期的一份研报中分析，国产超节点迎来量产元年，超节点架构通过系统工程扩展有效算力，算力竞争范式逐步切换。这也意味着，国产算力竞争正从单点峰值性能全面转向系统级全栈效率的比拼，集群工程能力与软件生态成为核心壁垒。 微信编辑| 夏木 ·联系我们 线索爆料：bianjibu@yicai.com 商务合作：business@yicai.com  阅读原文
Key Quotes
> We have always believed that the true explosion of the AI era will not come from one or two chips, but from a set of more open, coordinated, and efficient full-stack strategy computing systems.
> Open-sourcing the AI software stack is just the beginning; Pingtouge hopes to collaborate with global developers, research institutions, and industry partners to jointly refine toolchains, enrich the operator ecosystem, and optimize performance experience.
> A single cabinet integrates 128 chips, achieving Pb/s-level bandwidth and hundred-nanosecond-level latency through the self-developed ALink interconnect architecture.
Tags
Domestic Computing Power
AI Chip
Software Stack
Supernode
Open-source Ecosystem
Related Articles
* AI Arrival: From Technological Surge to Humanistic Awakening | Liang Shuo * Pre-planning for Elderly Care: They Hope to Age with Dignity * Conversation with NIO Senior Vice President Ren Shaoqing: How to Adapt One World Model to Two Chips, Four Platforms, and Over a Dozen Vehicles? * Japanese Corporate Bankruptcies Hit Record High * Beware the Risk of a Crash in the World's 'Largest' Chip IPO * Meituan Officially Releases LongCat-2.0: A Trillion-Parameter Model Fully Trained and Inferred on Domestic Computing Clusters * A Conversation with Chen Boyuan, a 2004-born World Model Entrepreneur: I'm Not a Genius * Jensen Huang's RTX Spark Makes Real Appearance at Bilibili World! CPU and GPU Directly Soldered, Laptop Runs 120B Large Model * Alibaba's Li Feifei Debut: 32 New Products for Agents in One Go * Hefei Sci-Tech Ecosystem Observation: Financial Collaboration Supporting Industrial Advancement.")