博通3nm芯片，重磅发布

Core Viewpoint - The article discusses Broadcom's new Jericho4 network chip, which addresses the challenges faced by cloud computing companies in managing aging and smaller data centers, particularly in the context of increasing demands for artificial intelligence (AI) workloads [2][5]. Group 1: Product Features and Capabilities - The Jericho4 chip can connect over 1 million processors across multiple data centers, processing information at four times the capacity of its predecessor [2][5]. - It supports distances over 100 kilometers, enabling the interconnection of geographically distributed data centers, which is essential for AI training and inference workloads [5][10]. - Each system can support up to 36,000 ports, with each port providing 3.2 Tbps bandwidth through Broadcom's proprietary HyperPort interface [7][9]. Group 2: Market Demand and Applications - Broadcom is benefiting from the growing demand for devices used in building AI systems, as its network components facilitate traffic between expensive GPUs used for AI model creation [2][3]. - The need for distributed AI computing is increasing, with companies looking to migrate data center capacity closer to customers to speed up AI model responses [3][5]. Group 3: Technical Innovations - HyperPort technology improves bandwidth efficiency and reduces congestion, achieving up to 70% better bandwidth utilization compared to traditional methods [10][11]. - Jericho4 integrates line-rate MACsec encryption on each port, ensuring secure traffic between facilities without impacting performance [13]. - The chip's architecture allows for seamless integration with existing Ethernet networks, simplifying deployment and reducing vendor lock-in for customers [13][14]. Group 4: Strategic Positioning - Jericho4 complements Broadcom's Tomahawk and Trident platforms, focusing on interconnectivity between facilities while maintaining the same management models and routing strategies [14][16]. - Broadcom's "Ethernet-first" strategy positions Jericho4 as a viable solution for meeting the demanding requirements of AI workloads, moving away from traditional architectures like Infiniband [16].