人工智能网络基础设施的未来走向

What’s Next for NetworkingInfrastructure for AI Sponsored by: Highlights ofWhat’sNext forNetworking Infra for AI •AIinfrastructure needs con0nue to expand.Enterprises are star+ng to adapt large language models(LLMs) to fit their specific businessrequirements.A mixture of infrastructurewill beneededto deliver •Inferencingneedswillexpandinfrastructurein a variety of ways.AI inferencing, whichenablesapplica+onsto takeinputand process output throughAI models,will be distributed acrosscloud andenterprise infrastructure.Asmodels evolve, moreinferencing infrastructure will be needed to interpret •Ethernet is gaining ground against InfiniBand.Efforts to shiH AI networks from reliance on NVIDIA’sproprietary InfiniBand networking technology is showing drama+c results, with adop+on of Ethernet •The Ultra Ethernet Consor0umremainsrelevant.Efforts by vendors,including NVIDIA,are coalescingaround a standard that improves on Ethernet’s drawbacks and RDMA’s limita+ons. •Speeds are increasing.While mostAI datacenter switches support speeds of 400-Gb/s, 800-Gb/s ratesare increasingly on the horizon, with even higher speeds in the works. •Op0cal networking is part of AI’s future.As AI networks grow in scale,speed, and power requirements,op+cal components will furnish solu+ons that save power, space, and opera+onal costs. •AIspecialized processors such as SmartNICs, IPUs, and DPUs are growing in importance for AIinfrastructure.Network interface cards powered by specialized chips are key to enabling beVerperformance of networking, security, and storage func+ons of AI networks. But switch vendors are •SASE, SD-WAN, andnetwork-as-a-servicewillevolve tosupport AInetworking with more pervasivesecurity.AI will increase data traffic by orders of magnitude, but adistributed networkinginfrastructure •Observability and AIOps are central to AI networking ROI.The ability to track, analyze, and automatenetworking efficiency is becoming vital to enterpriseadop+on of AI networking. •Companiesincluded in this report:Akamai,AMD, Arista, Arrcus,Aryaka,Astera Labs,Aviz NetworksAWS, Broadcom, Ciena, Cisco,Cloudflare,CoreWeave, DriveNets, Enfabrica,Equinix,Google Cloud,Hedgehog,Infinera, Juniper Networks, Lambda Labs,Marvell,Meta, MicrosoH,Napatech,Netris,NVIDIA,Oracle,Vapor IO,VersaNetworks,xAI, ZEDEDA Highest PerformanceEthernet Alternativeto InfiniBand DriveNets Network Cloud-AIdelivers industry-leading AInetworking fabric performance DriveNets Network Cloud-AI offers open networking infrastructure solutionsfor all sizes of GPU clusters and applications – from Enterprise to NeoCloudto Hyperscalers. Optimal solution for Enterprise applications including forHealthcare, Financial Services, Energy, Autonomous Driving and Defense. Table of Contents 1.Intro:AIBoosts theNetworking Market -From Founda*onal LLMs to the Edge-Security and Architecture Implica*ons 2.AI Networking in the Core: The Arrival of Ethernet -Scaling AI Clusters in the Core-The Arrival of Ethernet and the UEC-IncumbentNetworking Vendors Suppor*ngtheEthernet Trend-Startup NetworkingSolu*ons-Building Out Hyperscaler Stacks-GeQng Smart About NICs-What About Op*cs?- 3.Enterprise AI at the Edge -RAG and SLMs at the Edge-SD-WAN, SASE, and NaaS Adapt for AI-Greater Focus on MCN and Hybrid Networking-Orchestra*on at the Edge 4.Conclusion What’s Next for Network Infrafor AI|2025 1.Introduc,on: AIBooststhe Networking Market Ar+ficial intelligence (AI) heads thelist of corporate goals. Firms everywhere are kicking the +res of AI,engaging in proof-of-concepts, and strategizing to engage the necessary talent to bring genera+ve AI tobear on a range of cri+cal applica+ons. At the same +me, it’s clear that to reap AI’s rewards requires a In many ways, networking determines AI success.AI requires new levels of performance in bandwidth,latency, and security.Onlythe mostefficient and top-performing networks candeliver the data needsforinferencing,or the adapta+on of trained large language models (LLMs) such as ChatGPT, Llama, More importantly, AI is changing en+re network architectures. Enterprises now need to think aboutwhat networks are needed to support AI whether that’s in the core or at the edge. They also need tothink about what impact AI applica+ons will have on corporate networks, datacenters, and governancestrategies. If you are pucng your most valuable and proprietary data into AI models, you certainly need So far, the shiHs in networking for AI have fallen into two general categories: 1)Networkinginfrastructurefor AI, or the connec+vity infrastructure needed for AI datacenters and inference; and 2)AI for networking—the AI automa+on to drive opera+ons, oHen also referred to as AIOps. In this report, AI Democratization:From Foundational LLMs to the Edge This report covers how AI is impac+ng networking infrastructure from the core the edge, whether it’s adatacenter in SeaVle or an inference network in Dubai. So far, much of the aVen+on has been placed onthe hype