Performance profiling and troubleshooting: Practice Questions — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)

Performance Profiling and Troubleshooting Practice Questions

This set of multiple-choice questions is designed to help candidates prepare for the Performance Profiling and Troubleshooting portion of the GPU Acceleration and Optimization topic in the NVIDIA-Certified Professional: Generative AI LLMs exam.

1. Which NVIDIA tool is primarily used for detailed GPU kernel profiling to identify bottlenecks in large language model training?

   • A. Nsight Systems
   • B. Nsight Compute
   • C. CUDA-GDB
   • D. NVIDIA Visual Profiler

   Correct Answer: B

   Explanation: Nsight Compute provides detailed kernel-level profiling, allowing precise identification of GPU bottlenecks, which is essential for optimizing LLM training performance.

2. During distributed training, you observe that GPU utilization is low despite high CPU usage. What is the most likely cause?

   • A. Insufficient batch size causing underutilization
   • B. Network communication overhead between GPUs
   • C. Overlapping data transfers with computation
   • D. Excessive memory allocation on GPUs

   Correct Answer: B

   Explanation: High CPU usage combined with low GPU utilization often indicates communication overhead in multi-GPU setups, where data synchronization delays GPU compute.

3. Which metric is most useful to identify if kernel launches are serialized and causing performance degradation?

   • A. SM occupancy
   • B. Warp execution efficiency
   • C. Kernel launch latency
   • D. Memory throughput

   Correct Answer: C

   Explanation: Kernel launch latency measures delays between kernel executions; high latency suggests serialization that can reduce parallelism and throughput.

4. What is the first step in troubleshooting a sudden drop in GPU memory availability during training?

   • A. Increase batch size to utilize memory better
   • B. Profile memory allocation using tools like nvprof or Nsight
   • C. Restart the training job
   • D. Disable mixed precision training

   Correct Answer: B

   Explanation: Profiling memory allocation helps identify leaks or unexpected memory usage before making configuration changes or restarting.

5. Which of the following techniques helps isolate performance issues caused by inefficient data loading?

   • A. Using CUDA streams to overlap data transfer and computation
   • B. Increasing the number of GPUs
   • C. Reducing batch size
   • D. Disabling mixed precision

   Correct Answer: A

   Explanation: Overlapping data transfer and computation with CUDA streams reduces idle GPU time caused by slow data loading, improving throughput.

6. When profiling reveals low SM (Streaming Multiprocessor) occupancy, what is a common optimization to improve GPU utilization?

   • A. Increase the number of CUDA blocks per grid
   • B. Reduce batch size
   • C. Disable tensor cores
   • D. Use single precision instead of mixed precision

   Correct Answer: A

   Explanation: Increasing CUDA blocks per grid can improve occupancy by providing more parallel work to the GPU, thus better utilizing SMs.

7. In a multi-GPU setup, what profiling approach helps identify if workload imbalance is causing some GPUs to idle?

   • A. Profiling individual GPU kernel execution times
   • B. Monitoring inter-GPU communication bandwidth
   • C. Measuring per-GPU utilization and compute time
   • D. Checking batch size per GPU

   Correct Answer: C

   Explanation: Measuring per-GPU utilization and compute time reveals if some GPUs finish earlier and wait, indicating workload imbalance.

8. What is a common cause of performance degradation that can be detected by analyzing CUDA stream concurrency?

   • A. Excessive memory fragmentation
   • B. Serialization of kernel launches
   • C. Insufficient GPU memory
   • D. Incorrect batch normalization parameters

   Correct Answer: B

   Explanation: Lack of concurrency in CUDA streams leads to serialized kernel execution, reducing parallelism and overall GPU throughput.

More in this topic

Memory and batch optimization — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Parallelism techniques: Common Mistakes — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Parallelism techniques: Practice Questions — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Performance profiling and troubleshooting: Quick Reference — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Multi-GPU and distributed setups: Quick Reference — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Multi-GPU and distributed setups: Worked Example — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Parallelism techniques: Quick Reference — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Performance profiling and troubleshooting: Common Mistakes — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Multi-GPU and distributed setups — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Performance profiling and troubleshooting — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)GPU Acceleration and Optimization — NVIDIA-Certified Professional: Generative AI LLMsMulti-GPU and distributed setups: Common Mistakes — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Performance profiling and troubleshooting: Worked Example — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Parallelism techniques: Worked Example — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Parallelism techniques — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)Multi-GPU and distributed setups: Practice Questions — GPU Acceleration and Optimization (NVIDIA-Certified Professional: Generative AI LLMs)