Networking Service
Say goodbye to communication bottlenecks
Developing a sub-nanosecond AI compute network
Technological Advantages
Ultra-Low Latency Architecture
End-to-end latency as low as 0.8 µs (20x faster than traditional TCP/IP)
Supports 200/400/800 Gbps InfiniBand standards
Topology-aware Scheduling
Dynamic routing optimization and automatic inter-rack traffic balancing Deep optimization of the NCCL library results in a 40% acceleration of AllReduce operations
Service Benefits
Single-rack Topology
32 H100 GPUs fully interconnected via NVSwitch
900GB/s bisection bandwidth
Supports Multi-cluster Scaling
SHARP-based distributed routing
enabling expansion to 1,000-10,000 nodes
Solution Features for Accelerated Large Model Training
GPUDirect RDMA and NCCL topology optimization
Results: 175B parameter model training task
communication overhead was reduced from 35% to 8%, inter-rack traffic was decreased by 60%
Balanced Storage-Compute Convergence: The Three Pillars of Optimal Performance
Dimensions
- Bandwidth
- 低延时
- Scalability Unit
Network Layer
- 400Gbps
- 0.6μs
- 36-Port Lossless Fabric
Compute Layer
- 1.8TB/s NVLink
- 0.3μs GPU Interconnect
- 8-GPU Node
Storage Layer
- 320GB/s All-Flash
- 200μs Random Read
- 100PB Unified Namespace
