NVIDIA-Certified Professional: AI Infrastructure Practice Questions: Troubleshooting and Optimization Domain

Correct Answer: A

Explanation: CPU limitations are a common bottleneck in multi-GPU setups, as the CPU must coordinate tasks among GPUs. Insufficient RAM and network bandwidth affect data handling and transfer but are less likely to cause scaling issues directly. GPU driver issues typically cause errors rather than scaling problems.

Correct Answer: B

Explanation: Container GPU performance issues are often related to improper GPU device exposure. Using --gpus all with proper device isolation ensures containers have full access to GPU capabilities without interference from other containers.

Correct Answer: C

Explanation: nvprof is a profiling tool that provides detailed performance metrics, including memory bandwidth usage, which is crucial for identifying bottlenecks. nvidia-smi is more suited for high-level monitoring, CUDA-MEMCHECK is for debugging memory errors, and CUDA-GDB is for debugging at the code level.

Correct Answer: A

Explanation: Installing additional cooling units is the most direct way to address thermal throttling by reducing the temperature around the GPUs. Reducing workload (B) does not solve the underlying cooling issue. Increasing ambient temperature (C) would worsen the problem. Updating firmware (D) is unlikely to address thermal issues unless specifically noted in release notes.

Correct Answer: C

Explanation: Proper airflow direction and volume are crucial for effective cooling in high-performance systems like the NVIDIA DGX. Rack space availability and ambient room temperature are important but secondary to ensuring efficient airflow. Network cable management, while important for organization and troubleshooting, does not directly impact cooling.

Correct Answer: B

Explanation: Asynchronous kernel launches can help reduce latency by allowing the CPU to continue executing other tasks while waiting for the GPU to finish. Increasing threads per block (A) or GPU clock speed (D) does not address launch latency directly. Optimizing algorithms (C) can help but is not specifically targeting launch latency.

Correct Answer: A

Explanation: Recompiling with a higher optimization level can lead to more efficient code execution. Increasing threads or switching to double precision without understanding the workload may not be beneficial, and reducing kernels can decrease parallelism.

Correct Answer: D

Explanation: Dust buildup is a common cause of overheating and is relatively easy to address. Cleaning the GPU should be the initial step before considering more invasive measures like replacing thermal paste (A) or rearranging hardware (B). Increasing fan speed (C) might help temporarily but doesn't address the underlying issue.

Correct Answer: C

Explanation: InfiniBand switch configuration errors can lead to suboptimal routing and reduced communication speeds between nodes, affecting performance. Option A is incorrect as LinkX cables, if faulty, would typically cause connectivity issues rather than slow communication. Option B is incorrect because CUDA kernel configuration affects computation, not inter-node communication. Option D is incorrect as GPU memory affects computation capacity, not communication speed.

Correct Answer: B

Explanation: CUDA Profiler is designed to analyze CUDA applications and identify performance bottlenecks. nvidia-smi (A) provides GPU utilization metrics but not detailed application profiling. dmesg (C) is for kernel messages, not application profiling. top (D) shows CPU usage and process information but is not specific to CUDA.