NCA-GENL Software Development Practice Question
Exhibit
{
"model_name": "llama-3-8b",
"max_batch_size": 128,
"precision": "fp16",
"enable_cuda_graph": true
}Refer to the exhibit. An engineer is tuning a deployment config. Why is 'enable_cuda_graph' set to true in this JSON configuration?
⚠ Common exam trap
Candidates often assume CUDA Graphs are used for distributed multi-node communication or automatic mixed precision, missing their actual purpose of eliminating CPU launch overhead.
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
✓
To reduce CPU overhead during repetitive kernel launches.
CUDA Graphs capture a sequence of GPU work as a single graph, reducing CPU overhead associated with kernel launches. This is critical for LLMs where many small kernel calls can lead to CPU-bound execution. In high-performance generative AI scenarios, reducing launch latency is essential to ensure that the GPU remains saturated with work, thereby maximizing tokens-per-second and reducing total request latency for end-users.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
To increase the maximum batch size to 256.
Why it's wrong here
CUDA Graphs are unrelated to the batch size limit. The batch size is a memory-constrained parameter, whereas the graph feature is a runtime optimization designed to minimize CPU interaction by grouping GPU task submissions into a single, pre-compiled workload.
- ✓
To reduce CPU overhead during repetitive kernel launches.
Why this is correct
By capturing the graph of operations, the driver can execute them with a single launch command. This avoids the overhead of traversing the command queue for every operation, which is highly beneficial for LLM inference where the execution pattern is consistent.
- ✗
To convert the model precision from fp16 to fp8.
Why it's wrong here
Precision conversion is handled by the model compilation step or quantization tools, not by the CUDA Graph feature. The precision field in the JSON is static and independent of the execution graph optimization being enabled in the configuration.
- ✗
To force the model to use the CPU for inference.
Why it's wrong here
CUDA Graphs are strictly a GPU-side optimization. They cannot be used to offload inference to the CPU. In fact, they are designed to minimize the CPU's influence on the execution pipeline to ensure the GPU maintains peak operational efficiency.
About these practice questions
This NCA-GENL question is part of Courseiva's 367-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. 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.