Articles by tag "AI cluster networking"

1. Cost-Effective Deployment of NVIDIA Mellanox Switches for High-Bandwidth AI Clusters
   Selene Gong

   36
   High-bandwidth AI clusters rely on low-latency, high-throughput networking to support demanding workloads such as machine learning training, inference, and large-scale simulations. NVIDIA Mellanox switches deliver high port density, scalable throughput, and robust reliability. This guide ...
2. Best 100G Network Switches for AI Clusters in 2026: Architecture, Performance & Deployment Guide
   Selene Gong

   33
   As AI infrastructure rapidly scales toward thousands of GPUs per cluster, network design has become one of the most critical determinants of training efficiency. In 2026, 100G Ethernet switches are no longer optional—they are the baseline fabric for AI clusters, enabling high-throughput GPU ...
3. NVIDIA Mellanox MQM8790-HS2F Lead Time: How to Avoid Delivery Delays with Ready-to-Ship Inventory
   Selene Gong

   18
   Quick Take Navigating extended manufacturer backlog cycles for the NVIDIA Mellanox MQM8790-HS2F in 2026 requires shifting from factory allocation queues to ready-to-ship global inventory channels. Securing brand-new, original factory-sealed (Original Sealed) switches from centralized stock ...
4. InfiniBand vs RoCE v2 for 100B LLM Training: 2026 Architecture Guide
   Selene Gong

   18
   Quick Take For 100-billion parameter LLM training, 200G InfiniBand consistently outperforms 400G RoCE v2 in Model FLOPs Utilization (MFU). While RoCE v2 provides higher raw bandwidth at a lower upfront cost, its reliance on PFC and ECN introduces tail latency jitter and pause frames under ...
5. Quantum to Quantum-2: Can MQM8790 (HDR 200G) and MQM9790 (NDR 400G) Coexist in a Single AI Cluster?
   Selene Gong

   18
   Quick Take While the 200G HDR Mellanox MQM8790 (Quantum) and 400G NDR MQM9790 (Quantum-2) switches can logically coexist within the same data center operational environment, mixing them into a single, fully unified symmetric fabric is strongly discouraged. Cross-generation link speeds, ...
6. 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 ...
7. The Ultimate MQM8790 Cabling Guide: DAC vs AOC vs Optical Transceivers (1m–100m Matrix)
   Selene Gong

   18
   Quick Take At 200G/400G line rates, the physical layer is a core determinant of fabric stability. The optimal MQM8790 interconnect strategy is explicitly defined by distance: standardize on passive copper DACs for intra-rack links under 3 meters to minimize latency and power, ...
8. ConnectX-6 vs ConnectX-7: Best NIC Choice for MQM8790 200G InfiniBand Clusters (ROI Guide)
   Selene Gong

   18
   Quick Take While the MQM8790 switch delivers a non-blocking 200G InfiniBand fabric, server-edge throughput limits are ultimately dictated by the host NIC interface. ConnectX-6 utilizes PCIe Gen4 x16, which operates near saturation when handling sustained 200G line-rate traffic, whereas ...