NCA-AIIO Question 106
Select 4You are responsible for monitoring a cluster of GPUs in a data center. During a routine check, you notice that the performance of AI workloads has degraded. Which of the following are key metrics to monitor in order to diagnose potential GPU-related issues?
- A
GPU utilization percentage
- B
Memory temperature
- C
Disk read/write speed
- D
GPU memory usage
- E
PCIe bandwidth utilization
- F
CPU core temperature
Show answer and explanation
Correct answers: A, B, D, E
Explanation
Monitoring GPUs effectively requires focusing on metrics that directly impact their performance and operational health, such as utilization, memory usage, temperature, and data transfer throughput (e.g., PCIe bandwidth). These metrics help diagnose issues related to workload inefficiencies, overheating, and resource bottlenecks. Metrics like disk read/write speed and CPU temperature, while relevant to overall system performance, are not primary to GPU monitoring.
- A. Correct.
GPU utilization percentage is a key metric to monitor as it shows how effectively the GPU is being used. Low utilization during an AI workload might indicate bottlenecks in other parts of the system.
- B. Correct.
Memory temperature is crucial because overheating can lead to throttling or hardware failure, directly impacting GPU performance.
- C. Incorrect.
Disk read/write speed is not directly related to GPU monitoring, although it may affect data throughput for workloads. It is not a primary GPU monitoring metric.
- D. Correct.
GPU memory usage indicates whether the workload is exceeding the memory capacity of the GPU, which could result in performance degradation or errors.
- E. Correct.
PCIe bandwidth utilization is important for understanding if data transfer between the CPU and GPU is a bottleneck, which can impact workload efficiency.
- F. Incorrect.
CPU core temperature is not directly related to GPU performance, though overall system health might be influenced by it. It is not a key GPU monitoring metric.