Online Access Free NCP-AII Practice Test

A system administrator is working on a DGX A100 and needs to start a CUDA job. The job fails and a cudaErrorSystemNotReadyerror is returned. What is the cause of this error?

• A.There are too many running CUDA jobs.
• B.A GPU fails to register with the fabric.
• C.CUDA is not properly installed.
• D.FM is not completely done initializing.

An engineer needs to verify NVLink isolation on a single node with 8 GPUs. Which NCCL test configuration stresses switch bisection bandwidth?

• A.Use NCCL_TESTS_SPLIT= "DIV 8" with point-to-point tests
• B.Use reduce_scatter_pref -b 8 -e 16G -f2 -g 4
• C.Use all_reduce_perf -b 8 -e 16G -f2 -g 8 with NCCL_TESTS_SPLIT= "AND 0x1"
• D.Use all_reduce_perf -b 8 -e 16G -f2 -g 8 without splits

During HPL execution on a DGX cluster, the benchmark fails with "not enough memory" errors despite sufficient physical RAM. Which HPL.dat parameter adjustment is most effective?

• A.Reduce the problem size while maintaining the same block size.
• B.Increase block size to 6144 to maximize GPU utilization.
• C.Set PMAP to 1 to enable process mapping.
• D.Disable double-buffering via BCAST parameter.

You are an infrastructure engineer tasked with validating a new AI training cluster before releasing it to users. Your team wants to perform a NeMo burn-in to ensure both hardware and software are reliable and ready for production workloads. Which of the following actions are required as part of a proper NeMo burn-in process? (Choose two.)

• A.Download a pre-trained NeMo model and use it for a quick accuracy check on a user dataset, then consider the burn-in complete if results are reasonable.
• B.Run the configured NeMo training job repeatedly or for an extended duration, monitoring for errors, stalls, or performance drops across all GPUs and nodes.
• C.Test inference using the NeMo API and approve the environment if the model outputs valid predictions.
• D.Configure a representative NeMo training or pretraining recipe and set up an executor to launch the job across intended nodes and GPUs.

What is the primary purpose of running a NCCL burn-in test on a new GPU cluster?

• A.To test if GPUs are properly detected by the operating system and have the right drivers installed.
• B.To benchmark the application-specific runtime performance of AI models using real user data and production training scripts.
• C.To detect and resolve hardware or interconnect issues before production, by stressing GPU communication links.
• D.To maximize the GPU utilization for machine learning workloads and to automatically tune deep learning frameworks.