Telemetry Now From GPUs to Packets: The Critical Role of Networks in AI Data Centers
14 snips
Aug 29, 2025 
In this episode, Vijay Vusirikala, an industry leader at Arista with a rich background in large-scale optics and data centers, delves into the pivotal role networks play in AI data centers. He emphasizes that job completion time is the key performance metric, not just GPU power. Vijay discusses the significance of lossless networks, the challenges of optical power, and the importance of minimizing network tiers for efficiency. He also highlights emerging observability needs and the future of scalable inter-rack connectivity.
AI Snips
Chapters
Transcript
Episode notes
Network As The Critical Bottleneck
- The network is the limiting factor for GPU performance in AI training, not just the GPU count.
- Job completion time is the true metric that determines competitiveness and time-to-market.
Career Path Shapes AI Networking View
- Vijay described his career path from Google to sovereign cloud and now Arista, focused on energy-efficient AI data centers.
- He uses customer-facing systems work to align Arista's roadmap with hyperscaler needs.
AI Clusters Require Multiple High-Bandwidth Fabrics
- AI clusters add backend and scale-up networks that far exceed front-end bandwidth needs.
- These extra networks drive much higher power, capital expense, and cabling complexity.