AIInfoQabout 2 hours ago
Presentation: Chaos Engineering GPU Clusters
1 min read
Bryan Oliver presented seven fault-injection strategies for GPU clusters at QCon AI. An NVIDIA H100 has 94 GB VRAM and 16,000 CUDA cores. A GB200 Grace Blackwell rack costs approximately $3 million. XID errors, ECC, NVLink, and thermal throttling are key failure modes. DCGM from NVIDIA provides GPU management and observability tools.
Level
Hype check
Tap to vote and see what everyone thinks.
