NCA-ADS - Software and Environment Management (6% of the exam) - Section 8.3
Perform GPU environment checks.
Perform GPU environment checks using tools such as nvidia-smi and CUDA device queries to confirm driver version, device availability, and memory capacity before running workloads. Diagnose common configuration issues including driver/CUDA version mismatches and insufficient device memory.
Practice question for this objective
lock_openFree sampleSoftware and Environment Managementeasy
While checking a new environment, a developer runs nvidia-smi and reads a CUDA Version of 12.4 in the top-right header. They want to know what this particular number tells them about the system. Which interpretation is correct?
AIt is the highest CUDA runtime version the installed driver can support, not necessarily the toolkit version installed on the system.check_circle Correct
BIt is the version of cuDF currently installed, which determines which RAPIDS APIs are available in the environment.
CIt is the exact version of the CUDA toolkit installed for compilation, which nvcc would report identically on this machine.
DIt is the compute capability of the physical GPU, which fixes the set of CUDA features the hardware can execute.
Read the nvidia-smi CUDA Version header as the maximum CUDA runtime the driver supports, not the installed toolkit version. The CUDA Version shown by nvidia-smi reflects the highest CUDA runtime API the loaded driver can service, which is why it can exceed or differ from the toolkit version reported by nvcc; the driver remains backward compatible across older runtimes.
Why A is correct: The nvidia-smi header reports the maximum CUDA runtime API version the current driver is capable of supporting, so a separately installed toolkit may be the same or older, which is the precise meaning of that field.
Why B is wrong: This is tempting because RAPIDS versions track CUDA closely, but the nvidia-smi header reports nothing about cuDF or any Python package. Installed RAPIDS versions are shown by pip or conda, not by nvidia-smi.
Why C is wrong: This is a frequent confusion: the header shows the maximum CUDA runtime the driver supports, not the installed toolkit. nvcc reports the toolkit version, which can differ from and be older than the driver figure.
Why D is wrong: This is plausible because compute capability is a real GPU property, but it is expressed as a separate value such as 8.0 or 9.0 and is not what the CUDA Version header shows. The header is a driver-support figure.
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-ADS bank for this domain.
It will fit, because cuDF compresses tabular data so the 30 GB dataset occupies well under the 26 GB of free memory currently reported.
errorWhy it is wrong: This is tempting because Parquet on disk is compressed, but cuDF holds columns uncompressed in device memory, so the dataset needs roughly its full size plus working space, and assuming compression in VRAM is a common misconception.
Reinstall cuDF using a different channel, because the message indicates a package corruption that can be fixed only by a clean reinstall.
errorWhy it is wrong: Jumping to a reinstall before checking basic driver visibility is premature. The error is more likely caused by the driver not being accessible inside the environment than by a corrupt package, so diagnostic commands should come first.
Check that the cudf module imports without error, since a successful import guarantees the GPU is present and usable for computation.
errorWhy it is wrong: This is tempting because import failures are common, but importing cuDF can succeed while the GPU or driver is still misconfigured. The import does not by itself allocate device memory or confirm a usable GPU.