WZWeiwei Zhanginaicplight.hashnode.dev·Sep 3 · 2 min readWhat I Learned About Optical Networking While Studying AI Computing Center ArchitecturesOver the past year, I've spent a lot of time studying how AI infrastructure is evolving. One trend became increasingly clear: As AI clusters grow larger, networking is becoming a first-class design co00
WZWeiwei Zhanginaicplight.hashnode.dev·Sep 2 · 3 min readDesigning AI Clusters Beyond 1,000 GPUs: NDR or XDR InfiniBand?As AI models continue to grow, networking has become one of the most important factors affecting training efficiency. Most discussions around AI infrastructure focus on GPUs. However, once clusters ex00
WZWeiwei Zhanginaicplight.hashnode.dev·Aug 28 · 13 min readBuilding an InfiniBand Network for Large-Scale AI GPU ClustersAs AI models continue to grow in size and complexity, GPU clusters are becoming increasingly important to modern data center infrastructure. But adding more GPUs does not automatically guarantee bette00
WZWeiwei Zhanginaicplight.hashnode.dev·Aug 26 · 2 min readScale-Up vs Scale-Out vs Scale-Across: The Three Layers of Future AI InfrastructureAI infrastructure is evolving faster than ever. As model sizes continue growing, traditional scaling approaches are reaching practical limits. The industry is responding with a three-layer architectur00
WZWeiwei Zhanginaicplight.hashnode.dev·Aug 20 · 2 min readWhy AI Clusters Need New Optical Technologies Beyond 800GWhen people discuss AI infrastructure, GPUs usually get all the attention. But there is another critical component determining AI scalability: The network. As modern AI clusters grow larger, data move00
