Articles by tag "MQM8790"

1. Troubleshooting InfiniBand Subnet Manager Failover High-Availability OpenSM Architecture for MQM8790 Unmanaged Clusters
   Selene Gong

   18
   Quick Take In deployments based on unmanaged MQM8790 switches, the entire InfiniBand fabric control plane relies on an external Subnet Manager (SM). Implementing a high-availability active/standby OpenSM architecture with distinct sm_priority values prevents full cluster outages during ...
2. Maximizing MFU: How NVIDIA SHARP™ Inside MQM8790 Accelerates PyTorch Multi-Node Training
   Selene Gong

   18
   Quick Take NVIDIA SHARP™ inside the MQM8790 switch ASIC flattens the distributed training degradation curve by offloading collective AllReduce calculations from the GPU nodes directly into the network fabric. This in-network reduction reduces cross-tier traffic volume, minimizes ...