The demand for networks is surging in the AI era, and communication innovation is ushering in new opportunities.

robot
Abstract generation in progress

The Internet in the AI Era: Where Does Demand Come From and Where Will Innovation Head?

The network is a key component in the era of AI large models. In the era of large models, we have already seen accelerated iterations of network devices such as optical modules and switches, with an explosive demand. However, the market has a limited understanding of why graphics cards need to be paired with a large number of optical modules, and why communication has become a pain point for large models. This article will explore the principles behind why the network has become a new "C position" in the AI era, and discuss future innovations on the network side and the investment opportunities behind the latest industry changes.

Network demand mainly comes from three aspects:

  1. The increasingly large model size requires multi-server clusters to solve training issues.

  2. The network is more used to synchronize the model parameters between graphics cards, which places higher demands on network density and capacity.

  3. Large model training lasts for several months, the cost of network failures is extremely high, and the stability requirements are increased.

Network innovation is mainly reflected in three directions:

  1. Communication medium iteration: While optical modules pursue higher speeds, they also reduce costs through methods such as LPO, LRO, and silicon photonics. Copper cables dominate connections within cabinets. New technologies like Chiplet and Wafer-scaling accelerate the exploration of silicon-based interconnection limits.

  2. Network Protocol Competition: Communication protocols within nodes are bound to graphics cards, such as NVIDIA NV-LINK and AMD Infinity Fabric. The main competition between nodes is between IB and Ethernet.

  3. Changes in Network Architecture: The leaf-spine architecture is relatively costly in ultra-large clusters, and new architectures such as Dragonfly and rail-only architectures are expected to become the next generation evolution direction for ultra-large clusters.

Investment Advice:

Core links of communication systems: Zhongji Xuchuang, Xinyi Sheng, Tianfu Communication, Hudian Co., Ltd.

Innovation segments of the communication system: Yangtze Optical Fiber and Cable Joint Stock Limited Company, Zhongtian Technology, Hengtong Optic-Electric, and Shengke Communication.

Risk Warning: AI demand is below expectations, scaling laws have failed, and industry competition is intensifying.

ETH0.78%
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • 6
  • Share
Comment
0/400
MetaverseLandlordvip
· 13h ago
Investment opportunities have arrived, enter a position when the time is right.
View OriginalReply0
RugDocScientistvip
· 08-04 20:53
Is this wave of market purely relying on graphics cards?
View OriginalReply0
LadderToolGuyvip
· 08-04 20:50
Graphics cards have all risen crazily, who will take care of the internet fees?
View OriginalReply0
TheShibaWhisperervip
· 08-04 20:49
Focus on increasing the position in optical modules this wave~
View OriginalReply0
RugDocDetectivevip
· 08-04 20:48
The soaring graphics cards will eventually hit the wall of internet speed.
View OriginalReply0
SilentObservervip
· 08-04 20:33
The pace of the sales event is too fast, isn't it?
View OriginalReply0
Trade Crypto Anywhere Anytime
qrCode
Scan to download Gate app
Community
English
  • 简体中文
  • English
  • Tiếng Việt
  • 繁體中文
  • Español
  • Русский
  • Français (Afrique)
  • Português (Portugal)
  • Bahasa Indonesia
  • 日本語
  • بالعربية
  • Українська
  • Português (Brasil)