WWeiwei ZhangOct 13 · 6 min readHow InfiniBand XDR and 800G RoCE Handle Congestion in AI ClustersHigh-speed links are only one part of a high-performance AI network. In distributed training, thousands of GPUs may exchange data during collective operations such as All-Reduce. When multiple flows caicplight.hashnode.dev
NNathanael FETUEOct 7 · 6 min readMy RoCE testbed passed. It was lyingWhy I'm building this When a server in a GPU cluster dies, everyone knows within seconds. The failures that hurt more are the quiet ones: a link that still works but has become a bit slower, a NIC tha1Npegasus-nathanael-tech.hashnode.dev
WWeiwei ZhangAug 18 · 6 min readRoCE Explained: How RDMA over Ethernet Enables Scalable AI InferenceAs AI inference workloads scale across hundreds or thousands of GPUs, networking can become a bottleneck long before compute capacity is exhausted. A traditional TCP/IP network introduces CPU overheadaicplight.hashnode.dev
Aania0-artJul 26 · 9 min read0: Let's enjoy reading the paper!Almost all high-performance communication libraries rely on UCX at their core. A close reading of its foundational paper, "UCX: An Open Source Framework for HPC Network APIs and Beyond," published at once2twice.hashnode.dev
AAli MuJan 5 · 8 min readDetailed Explanation of RoCEv2: Technical Principles, Optimization Strategies, and Future TrendsIn modern data centers and high-performance computing (HPC) environments, the efficiency of network communication directly dictates overall computing and storage performance. Remote Direct Memory Access (RDMA) technology has emerged as a transformati...alisemiconductors.hashnode.dev
