share_log

国盛证券:AI以太网价值几何?

Guosheng Securities: What is the value of AI Ethernet?

Zhitong Finance ·  Jun 2 08:21

Guosheng Securities believes that the LPO solution may be the best option for Nvidia's AI Ethernet.

The Zhitong Finance App learned that Guosheng Securities released a research report saying that Nvidia (NVDA.US) first launched SpectrumX-800 Ethernet in March, which is suitable for AI-optimized networks for various data centers, and then indicated in the May financial report that the Specturm Ethernet platform (including DPU cards and switches) will contribute several billion dollars in revenue in the next year, all showing that Nvidia is actively promoting the spread of its network connections in the direction of Ethernet.

Guosheng Securities believes that the LPO solution may be the best option for Nvidia's AI Ethernet. LPO (linear drive pluggable optics, linear drive pluggable optical module) is mainly used in the field of high-speed optical modules. It replaces traditional DSP through linear direct drive technology, and has the advantages of low power consumption, low latency, low cost, and hot pluggability. However, since traditional DSP chips have been removed, the LPO solution requires cooperation with the switch. One of the biggest advantages of the Nvidia/UEC AI Ethernet solution over the IB solution is that it is lower in cost. In order to continue to implement the low cost and low power characteristics of the system, the LPO penetration rate is expected to accelerate in the future, but since this solution requires cooperation with switches, optical module manufacturers require higher upstream and downstream cooperation and collaboration within the industry, leading companies such as Zhongji Xuchuang (300308.SZ) and Xinyisheng (300502.SZ) will have more advantages.

Guosheng Securities's main views are as follows:

[Internet connection]

Marginal change: Nvidia “actively embraces” Ethernet. Nvidia first launched SpectrumX-800 Ethernet in March, which is suitable for AI-optimized networks in various data centers. The May financial report indicated that the Specturm Ethernet platform (including DPU cards and switches, etc.) will contribute several billion dollars in revenue in the next year, all showing that Nvidia is actively promoting the spread of its network connections in the direction of Ethernet.

Historical evolution: the “dispute” between Ethernet and IB networks.

General VS dedicated: Ethernet (Ethernet) is a local area network that is currently widely used. It uses standard Ethernet devices, cables, and interface cards to achieve communication between network nodes; IB networks (InfiniBand) are networks suitable for high-performance computing and AI scenarios, and use switches to achieve direct connection and efficient communication through special equipment, cables, and interface cards.

The original design led to a different focus on performance: the “dispute” between Ethernet and IB networks has been around for a long time. Initially, their design goals and application scenarios were different, which directly led to differences in performance focus. Overall, Ethernet has higher compatibility, lower cost, better flexibility, better IB network performance, and has the advantages of high bandwidth, low latency, high reliability, and high scalability.

Ethernet has a strong public base, and the IB network is backed by Nvidia's strong voice: Overall, as the most widely used communication protocol in the past few decades, Ethernet has a large audience, high user acceptance, lower cost, and strong compatibility. However, in the early stages of AI development, because it was rarely used in high-performance AI computing scenarios and scattered suppliers, the overall effect of Ethernet was lacking compared to closed Nvidia solutions. With its leading graphics card performance and user stickiness, Nvidia has made the IB protocol the top in the current AI connection field, but due to IB's high premium, users often need to pay more hidden networking costs. Under the same budget, the number of optical modules and switch hardware purchased for video cards will be smaller.

Looking to the end: Ethernet is the answer.

Nvidia ≠ IB Network: The main force for global cross-server AI connections is the IB protocol and the Ethernet protocol. The leading party of the IB network agreement is Nvidia, but this does not mean that Nvidia only accepts the IB solution. Judging from Nvidia's move to acquire Mellanox (an Ethernet switch and network interface card supplier) after acquiring InfiniBand, it can be seen that its long-term layout takes into account the Ethernet solution.

The advantages of low cost and high performance of Ethernet are gradually becoming obvious: in Meta tests, the price of Ethernet networking is less than half that of IB, and the performance is 10% higher, confirming that Ethernet buys more saves more.

Compared with traditional Ethernet, AI Ethernet tests the pain points of AI scenarios such as high performance, high stability, high bandwidth, and low jitter. Currently, the Super Ethernet Alliance and the Nvidia Spectrum-X platform have initially formed two mainstream camps:

(1) Super Ethernet Alliance UEC: It brings together equipment vendors such as AMD, Arista, Broadcom, and Cisco, and cloud vendors such as Meta and Microsoft. On the basis of RoCE, the advantages of combining Ethernet are more considered to be non-destructive networks to meet the requirements of AI and HPC in an open ecosystem.

(2) Nvidia Spectrum-X platform: A comprehensive system network architecture including software and hardware. Compared with traditional Ethernet, it has twice the AI cluster performance. The hardware-level product is the Spectrum-4 Ethernet switch (a key component of the Spectrum-X network). At the same time, the Spectrum-4 Ethernet switch is also connected to the BlueField-3 DPU, thereby improving node data transmission capabilities.

The bank believes that Ethernet has better performance, better large-scale clustering capabilities, and more core openness. As the subsequent AIGC development process accelerates, the demand for cluster capacity and computing power is even greater, and Ethernet is expected to gradually become the mainstream network interconnection solution. The dispute between IB and Ether may gradually fade out. At that time, the competition will exist between the Super Ethernet Alliance and Nvidia's AI Ethernet solutions. Whichever solution dominates the mainstream, it will promote the broadening of the switch market space and the acceleration of the overall development of the industry. Therefore, the bank suggests focusing on the switch industry. Chips, manufacturers, and brands are all expected to benefit from the gradual increase in the penetration rate of Ethernet solutions in AI application scenarios.

[Optical module]

The bank believes that the LPO solution may be the best option for Nvidia's AI Ethernet. LPO (linear drive pluggable optics, linear drive pluggable optical module) is mainly used in the field of high-speed optical modules. It replaces traditional DSP through linear direct drive technology, and has the advantages of low power consumption, low latency, low cost, and hot pluggability. However, since traditional DSP chips have been removed, the LPO solution requires cooperation with the switch. The bank believes that one of the biggest advantages of the Nvidia/UEC AI Ethernet solution over the IB solution is that it is lower in cost. In order to continue to implement the low cost and low power characteristics of the system, it is also preferred to use a low-cost LPO solution in the selection of optical modules, so the subsequent LPO penetration rate is expected to accelerate, but since this solution requires cooperation with switches, the requirements for upstream and downstream cooperation between optical module manufacturers are higher, and leading companies such as Zhongji Xuchuang and Xinyisheng will have more advantages.

Arithmetic power

Optical Communications: Zhongji Xuchuang, Xinyisheng, Tianfu Communications, Taichenguang, Tengjing Technology, Optical Bank Technology, Guangxun Technology, Decolly, Liante Technology, Huagong Technology, Yuanjie Technology, Cambridge Technology, Mingpu Optoelectronics. Computing power tuning/scheduling/leasing: Hengwei Technology, Startech, Zhongke Shuguang, China Mobile, China Unicom, China Telecom. Computing power equipment: ZTE, Ziguang Co., Ltd., Ruijie Networks, Shengke Communications, Phillincoth, IFF, Shanghai Electric Power Co., Ltd., Cambrian. Liquid cooling: Invec, Shenling Environmental, Gaolan Co., Ltd., Jialitu. Edge computing power carrying platform: MeiG Intelligence, Guanghetong, Yiyuan Communications. Satellite communications: China Satellite Communications, China Satellite, Zhenyou Technology, Huali Chuangtong, Electronics Chips, Haige Communications.

data elements

Operators: China Telecom, China Mobile, China Unicom. Data visualization: vast depth, Hengwei Technology, Sino-Singapore SEC. BOSS system: ASIX Technology, Tianyuan Deco, Dongfang Guoxin.

Risk warning: AI development falls short of expectations, computing power demand falls short of expectations, market competition risks.

Disclaimer: This content is for informational and educational purposes only and does not constitute a recommendation or endorsement of any specific investment or investment strategy. Read more
    Write a comment