You are assisting in a project that involves deploying a large-scale AI model on a multi-GPU server. The server is experiencing unexpected…

Analyzing GPU memory usage with nvidia-smi is the most effective approach to identify performance

degradation during inference on a multi-GPU server. NVIDIA’s nvidia-smi tool provides real-time

insights into GPU utilization, memory usage, and process activity, pinpointing issues like memory

overflows, underutilization, or contention—common causes of inference slowdowns. Option A

(power supply) is secondary, as power issues typically cause failures, not gradual degradation. Option

B (CPU utilization) is relevant but less critical for GPU-bound inference tasks. Option D (training data)

affects model quality, not runtime performance. NVIDIA’s performance troubleshooting guides

recommend nvidia-smi as a primary diagnostic tool for GPU-based workloads.

Reference: NVIDIA Developer Tools (developer.nvidia.com), Modern Data Centers (www.nvidia.com).

Below is the second batch of 10 questions (11–20) formatted as requested, with 100% verified

answers based on official NVIDIA AI Infrastructure and Operations documentation where applicable.

Each question includes a full, detailed explanation with

to NVIDIA resources at the end.

Typing errors have been corrected for clarity, and the answers align with the "Correct answer"

indications provided in your input.