近日，申万宏源黄忠煌团队发布研报称，在大模型参数呈爆炸式增长的当下，算力需求正从单点向系统 级整合加速转变。 那么，服务器厂商生存空间是否受到挤压？ 首先，AI 芯片厂商不会切入代工业务。AMD 收购 ZT System 后剥离了其代工业务，避免与 OEM/ODM 的竞争，海光收购曙光目的也是为了强化协同，提升液冷、软件等能力。 但是，算力链条的产业链分工可能会进一步细化。在超节点趋势下，AI 芯片之间、AI芯片与交换机芯 片之间的互联，大都需要通过板卡(尤其是电信号互联)。以英伟达为例，其板卡在产品推出初期自行设 计，产品稳定后会开放给 OEM 合作伙伴，此时板卡设计的能力就成为了能否获取更多价值量的核心差 异化能力。因此代工产业链分工可能进一步分化为板卡设计代工供应商、以及机柜代工供应商。 超节点实际就是算力网络系统在单个或多个机柜层面的 Scale up，节点内主流通信方案是铜连接与电气 信号，跨机柜则考虑引入光通信；其与 Scale out 的硬件边界是 NIC网卡，外部借助光模块、以太网交 换机等设备。二者的架构设计、硬件设备、协议标准有本质不同。 目前，Scale up 和 Scale out 尚 ...

如果您希望可以时常见面，欢迎标星收藏哦~ 来源：内容来自半导体行业观察综合 。 在本文中，我们来谈一下GPU集群的横向和综合拓展。 让我们从"AI Pod"的概念开始。这个术语对不同的人可能意味着不同的东西，但它通常指的是一种预先配置的模块化基础设施解决方案，旨在简化 和加速AI工作负载的部署。 这些"pod"将计算、存储、网络和软件组件集成为一个紧密相连的单元，从而促进高效的 AI 运行。这就是我们遇到"纵向扩展"和"横向扩展"等术语 的地方。以下是可视化示例： 每个 XPU 刀片通常包含 2 到 8 个 XPU 设备。每个设备可以形成为单片芯片（即由单个半导体切片制成），也可以形成由一组称为"芯片集"的芯 片组成的多芯片系统。 我们这里讨论的计算处理能力令人难以置信，XPU 设备本身也同样如此。例如，NVIDIA 的 B200 GPU 拥有超过 2000 亿个晶体管（当然，我可 没亲自数过）。 CPU（中央处理器） GPU（图形处理单元 NPU（神经处理单元） TPU（张量处理单元） DPU（数据处理单元） FPGA（现场可编程门阵列） ASIC（专用集成电路） END 半导体精品公众号推荐 扩展是有限制 ...

Summary of Astera Labs Conference Call (May 20, 2025) Company Overview - **Company**: Astera Labs (ALAB) - **Industry**: Semiconductor, specifically focusing on AI infrastructure connectivity solutions Key Points and Arguments 1. **AI Compute Demand**: The complexity of AI compute models is increasing, leading to larger AI compute clusters that require enhanced interconnectivity solutions [4][28] 2. **UA Link Technology**: UA Link is introduced as an open, high bandwidth, low latency connectivity architecture designed to improve interconnect fabric in GPU or XPU rack scale architectures [5][12] 3. **Product Portfolio**: Astera Labs has expanded its product offerings to include smart fabric switches, Ethernet retimers, and CXL controllers, all aimed at enhancing AI connectivity [14][16] 4. **Collaboration with NVIDIA**: Astera Labs is part of NVIDIA's NVLink fusion ecosystem, which aims to integrate custom compute and AI processing XPU systems [16][68] 5. **Market Opportunity**: The total addressable market (TAM) for Astera's Scorpio SmartFabric Switch family is estimated at approximately $5 billion by 2028, with UA Link expected to unlock additional multibillion-dollar opportunities [74][78] 6. **Challenges in AI Infrastructure**: Key challenges include power consumption, cluster utilization, and the integration of specialized AI accelerators into cloud infrastructure [31][32] 7. **Scalability and Efficiency**: UA Link aims to provide a scalable, efficient, and open connectivity solution that addresses the challenges of large cluster scaling and enhances total cost of ownership (TCO) [54][55] 8. **Interoperability**: The UA Link consortium is focused on creating an open, interoperable standard for XPUs, which will facilitate a resilient supply chain and enable multiple vendors to offer compatible solutions [56][57] Additional Important Content 1. **Memory Semantics**: UA Link utilizes a memory semantic protocol that simplifies the access mechanism for memory transactions across XPUs, enhancing efficiency [45][48] 2. **Switching Architecture**: The design of UA Link's switching architecture is kept simple to optimize performance and maintain low latency [49] 3. **Ecosystem Development**: The consortium is working on specifications for IO chiplets and in-network compute capabilities to further enhance the scalability and efficiency of AI infrastructure [50][83] 4. **Management Software**: Effective management software is critical for the integration and operation of the UA Link ecosystem, providing telemetry and cluster utilization information [63] 5. **Future Vision**: Astera Labs aims to be a leading supplier of connectivity solutions for AI at rack scale, continuously expanding its product lines to meet the evolving needs of the market [66][71] This summary encapsulates the core discussions and insights shared during the Astera Labs conference call, highlighting the company's strategic direction and the broader implications for the AI infrastructure market.