Return case listAI Data Center
High-performance network construction for AI computing centers
Build a low-latency, high-throughput RDMA network for AI training clusters to support large-scale model training workloads.
Project Background
AI large model training requires massive parameter synchronization across GPU clusters, demanding extremely low network latency and high throughput. Customers need to build a lossless RDMA network supporting RoCE v2 to ensure efficient inter-GPU communication.
Solution
Deploy CS9935Y series as Spine core and D7726C series as Leaf access to build a non-blocking Clos network. Enable PFC/ECN congestion control and RoCE v2 for low-latency, high-throughput data transfer between GPU clusters.
Use Product
CS9935Y SeriesD7726C Series
Project Outcomes
Reduced GPU intercommunication latency to microseconds
RDMA network throughput utilization exceeds 95%
30% boost in large model training efficiency
