NCA-AIIO Question 108
Select 3You are responsible for maintaining a cluster of NVIDIA GPUs used for AI training workloads. During routine monitoring, you observe performance degradation in one of the GPUs. Which of the following metrics or criteria should you prioritize to diagnose the issue?
- A
GPU memory utilization
- B
GPU temperature
- C
CPU utilization
- D
GPU power consumption
- E
Disk I/O speed
Show answer and explanation
Correct answers: A, B, D
Explanation
To diagnose GPU performance degradation, metrics directly related to the GPU's operation, such as memory utilization, temperature, and power consumption, are critical to analyze. These metrics provide insights into potential bottlenecks, thermal issues, or inefficient power usage. While CPU utilization and Disk I/O speed can impact overall system performance, they do not directly address GPU-specific performance issues.
- A. Correct.
Monitoring GPU memory utilization is critical to ensure that the workload fits within the GPU’s memory capacity. High memory utilization can lead to performance bottlenecks or memory-related errors.
- B. Correct.
GPU temperature is an essential metric to monitor. Excessive heat can lead to thermal throttling, which reduces GPU performance to prevent damage.
- C. Incorrect.
While CPU utilization is important for overall system performance, it is not a primary metric for GPU performance issues. It would be secondary unless the GPU workload is CPU-bound.
- D. Correct.
GPU power consumption is a key factor, as abnormal power usage can indicate hardware issues or inefficient workload execution.
- E. Incorrect.
Disk I/O speed is unrelated to GPU-specific performance monitoring and is typically relevant for data transfer bottlenecks or storage issues.