The rise of AI large models has led to an explosive demand for network devices, becoming a hot investment topic.

The Importance of the Internet in the Era of AI Large Models

In the era of large models, the demand for network devices such as optical modules and switches has exploded, accelerating iterations. This article will explore why networking has become a key link in the AI era and discuss innovations and investment opportunities on the network side.

Source of Network Demand

Entering the era of large models, the gap between model size and the single card limit is rapidly widening, leading the industry to turn to multi-server clusters to solve training problems. This forms the basis for the increased importance of networks in the AI era. Compared to the past, where data was simply transmitted, networks are now more used for synchronizing model parameters between graphics cards, requiring higher network density and capacity.

The increasingly large model size has led to an increase in training time. To shorten the training duration, it is necessary to improve computational efficiency, and expanding the "number of devices" and "parallel efficiency" directly determines the computing power.

In large model training, alignment is required between single cards after each computation, which places higher demands on network transmission and exchange.

Large model training often lasts for months, and interruptions can lead to huge losses. A failure or excessive delay in any part of the network can cause interruptions. Modern AI networks have become complex system engineering comparable to airplanes and aircraft carriers.

Network Innovation Direction

As the scale of computing power investment expands to hundreds of billions of dollars, cost reduction, openness, and balancing computing power scale have become the main topics of network innovation.

Medium Replacement in Communication: While optical modules pursue higher speeds, they also reduce costs through methods like LPO, LRO, and silicon photonics. Copper cables dominate cabinet connections due to advantages such as cost-effectiveness. New technologies like Chiplet and Wafer-scaling accelerate the exploration of the limits of silicon-based interconnects.

Network protocol competition: inter-chip communication protocols and strong bindings with graphics cards, such as NVIDIA's NV-LINK and AMD's Infinity Fabric. The competition between IB and Ethernet is the main theme of inter-node communication.

Changes in network architecture: Currently, the leaf-spine architecture is commonly used, but as the number of nodes increases, its redundancy characteristics lead to significant network costs for ultra-large clusters. The Dragonfly architecture, rail-only architecture, and others are expected to become the evolutionary direction for the next generation of ultra-large clusters.

Investment Advice

Core links of the communication system: Zhongji Xuchuang, Newyi Sheng, Tianfu Communication, Industrial Fulian, Yingweike, Hudian Co., Ltd.

Innovation links in communication systems: Changfei Optical Fiber, Taicheng Optoelectronics, Yuanjie Technology, Shengkai Communication-U, Cambrian, Dekoli.

Risk Warning

AI demand is below expectations, scaling laws have failed, and industry competition has intensified.