試験NCA-AIIO トピック2 問題3 スレッド
NVIDIA NCA-AIIOのリアル試験問題集
問題 #: 3
トピック #: 2
問題 #: 3
トピック #: 2
Your AI team notices that the training jobs on your NVIDIA GPU cluster are taking longer than expected.
Upon investigation, you suspect underutilization of the GPUs. Which monitoring metric is the most critical to determine if the GPUs are being underutilized?
Upon investigation, you suspect underutilization of the GPUs. Which monitoring metric is the most critical to determine if the GPUs are being underutilized?
おすすめの解答：A 解答を投票する
GPU Utilization Percentage is the most direct metric to assess whether GPUs are underutilized during training. Measured as a percentage of time the GPU is actively processing tasks, it's available via NVIDIA tools like nvidia-smi and DCGM (Data Center GPU Manager). A low percentage (e.g., below 70-80% during training) indicates the GPU isn't fully engaged, often due to bottlenecks like slow data loading or inefficient parallelism, common issues in NVIDIA GPU clusters (e.g., DGX systems). This metric pinpoints the root cause of prolonged training times.
Memory Bandwidth Utilization (Option B) shows memory usage efficiency but not overall GPU activity.
Network Latency (Option C) affects multi-node setups but isn't a primary indicator of single-GPU utilization.
CPU Utilization (Option D) reflects CPU load, not GPU performance. NVIDIA's performance tuning guides prioritize GPU Utilization for diagnosing underutilization.
Memory Bandwidth Utilization (Option B) shows memory usage efficiency but not overall GPU activity.
Network Latency (Option C) affects multi-node setups but isn't a primary indicator of single-GPU utilization.
CPU Utilization (Option D) reflects CPU load, not GPU performance. NVIDIA's performance tuning guides prioritize GPU Utilization for diagnosing underutilization.
Asaka 2026-03-07 11:18:52
コメント
他人の解答コメントを賛成するのも、その解答に一票を入れることになります。したがって、すでに同じ意見の投票コメントが存在する場合、新規コメントをする代わりに賛成することもできます。
コメントを通報する
コメント中
今すぐ 新規登録 / ログイン （無料です）。