NCA-GENL Core Machine Learning and AI Knowledge Practice Question
Refer to the exhibit. A machine learning engineer reviews the monitoring output from a two-GPU distributed training job running on an NVIDIA DGX system. GPU 0 shows low utilization despite high memory consumption, while GPU 1 shows high utilization and high memory consumption. What is the most likely root cause of this performance imbalance?
⚠ Common exam trap
Candidates often assume a hardware failure or driver mismatch when GPU utilization diverges. However, in distributed training, high memory with low utilization almost always indicates pipeline stalls, gradient synchronization bottlenecks, or uneven data loader chunking.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
Imbalanced data distribution or pipeline synchronization stalls causing GPU 0 to wait idly while holding allocated tensors.
Data parallelism with uneven batch distribution or pipeline parallel bubble inefficiency can cause worker synchronization stalls. When one GPU finishes its workload or waits for synchronization barriers while processing uneven tensor sizes or asynchronous communication primitives, its utilization drops while memory remains allocated to tensors.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
A PCIe bus bottleneck causing continuous memory swapping between host system RAM and GPU 0 device memory.
Why it's wrong here
PCIe swapping would show host-device transfer saturation and falling GPU memory residency, whereas GPU 0 shows high memory consumption, indicating data is resident. It tempts because PCIe bottlenecks do slow training, but the observed pattern reflects uneven work assignment rather than host-memory paging.
- ✓
Imbalanced data distribution or pipeline synchronization stalls causing GPU 0 to wait idly while holding allocated tensors.
Why this is correct
GPU memory remains fully allocated by framework tensors and optimizer states even when the compute units are idle. An imbalance in batch sizes or pipeline parallelism stages forces GPU 0 to wait at synchronization barriers, driving down its utilization metric.
- ✗
A hardware fault in the NVIDIA NVLink bridge connecting GPU 0 to the shared system NVSwitch fabric.
Why it's wrong here
An NVLink or NVSwitch fault would degrade inter-GPU communication and usually raise errors or stall both GPUs, not leave GPU 0 holding memory while idle. It tempts because NVLink carries gradient traffic in distributed training, but the symptom points to workload imbalance, not fabric failure.
- ✗
An incorrect CUDA driver version mismatch preventing GPU 0 from initializing its primary Tensor Core execution units.
Why it's wrong here
A driver mismatch would typically prevent initialisation or crash the job outright, not leave GPU 0 resident with high memory but idle compute. It tempts because version mismatches are a common CUDA failure, yet they manifest as errors rather than asymmetric utilisation across a healthy peer GPU.
About these practice questions
One of 367 original NCA-GENL practice questions on Courseiva, each with a full explanation and wrong-answer analysis — not exam dumps or protected exam content. Learn why practice questions differ from exam dumps →
JA
Written and reviewed by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
Last reviewed September 2026 · checked against the official NVIDIA exam blueprint
This NCA-GENL practice question is part of Courseiva's free NVIDIA certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the NCA-GENL exam.