公众号记得加星标⭐️，第一时间看推送不会错过。 数据中心通常被认为是 CPU（x86、ARM、RISC-V）和 GPU（NVIDIA、AMD、定制 ASIC）之 间的较量。但在这些备受瞩目的竞争背后，另一场静悄悄的革命正在上演：ARM 悄然取代了英特尔 和 AMD 在数据处理单元 (DPU) 市场的份额。 DPU（也称为 SmartNIC）负责数据中心的"管道"。它们通过管理数据包处理、TCP/IP 和 RDMA 来 减轻网络负载。它们处理压缩、加密和 NVMe-over-Fabrics (NVMe-oF) 等存储服务。它们强制执行 安全隔离，这在多租户云环境中至关重要，因为信任边界会不断受到考验。此外，它们还负责执行编 排任务，否则这些任务会浪费宝贵的 CPU 周期。 NVIDIA （ 通 过 Mellanox 开 发 的 BlueField ） 、 Marvell （ OCTEON ） 、 AMD （ Pensando ） 和 Broadcom 均在其 DPU 中采用了 ARM 内核。原因很简单：ARM 内核体积小、功耗低、可授权， 并且已经嵌入到网络芯片中。当英特尔推出其基础设施处理单元 (IPU) 计 ...

Core Viewpoint - Cisco is positioning itself as a leader in the AI era by introducing a new network architecture centered around its Silicon One chip, emphasizing programmability and embedded security as key differentiators in the competitive landscape of IT infrastructure [1][2]. Group 1: Silicon One Chip - The Silicon One chip is designed to meet the demands of AI workloads, addressing challenges such as routing logic, scalability, and security integration [1]. - Cisco has a long history of developing ASIC chips, with performance improvements over the past 20 years reaching 10,000 times [2]. - The chip's programmability allows it to adapt to new workloads without the need for re-siliconization, providing a scalable solution for enterprises [1][5]. Group 2: Collaboration and Ecosystem - Cisco's strategy includes a collaborative model where it offers various options for customers, including selling systems with Cisco software or providing chips for custom builds [4]. - The company recognizes the need for alternative suppliers to compete effectively in terms of product breadth and performance [4]. Group 3: Programmability and Flexibility - Programmability is highlighted as a crucial feature, enabling the network to support a wide range of applications, including AI agents [5][8]. - The flexibility of the Silicon One architecture allows for efficient traffic load balancing and the integration of security features into the network [6][14]. - Cisco aims to provide long-term investment protection for clients by ensuring that their solutions can adapt to future demands without requiring complete overhauls [9]. Group 4: AI Workloads and Future Developments - The focus is shifting from training AI models to inference, which will drive demand for more efficient and cost-effective computing solutions [10]. - Cisco plans to continue innovating within the Silicon One framework, potentially offering different versions of the chip to meet varying enterprise needs [12]. - The company is also exploring the integration of security features directly into the chip, enhancing the overall value proposition in the cybersecurity landscape [13][14].