NVIDIA Certified Associate - AI Infrastructure and Operations practice questions

Question 1

A customer is evaluating CUDA to accelerate a compute-intensive simulation that currently runs on CPUs and asks what CUDA fundamentally provides. Which statement matches the CUDA documentation?

Accepted Answer

It enables dramatic increases in computing performance by harnessing the power of the GPU.

Answer

It is a job scheduler that queues CPU and GPU tasks but performs no actual computation itself.

Answer

It converts CPU binaries to GPU code automatically with no changes to the application source.

Answer

It provides remote storage acceleration by moving files directly into GPU memory over a network.

Question 2

A team has decided to plan compute sizing separately for training versus inference workflows. Which set of characteristics matches the documented training profile on NVIDIA platforms?

Accepted Answer

Training demands maximum memory bandwidth (HBM3 at 3+ TB/s), high-precision floating point (FP32/TF32/BF16), massive parallelism across GPU clusters via NVLink and InfiniBand, and can run for days or weeks.

Answer

Training tolerates the lowest precision available (INT4/INT8) and ignores memory bandwidth entirely because gradients are computed only on CPU.

Answer

Training is always finished in seconds on a single GPU regardless of model size and never benefits from cluster-level interconnects like NVLink or InfiniBand.

Answer

Training is identical in compute profile to inference; the only difference is the file system path the model is loaded from on the host machine.

Question 3

An MLOps team is being built around high-throughput data loading and must choose between statements describing GPUDirect Storage. Select the TWO that match the documentation.

Accepted Answer

GDS creates a direct data path between local or remote storage, like NVMe-oF, and GPU memory.

Accepted Answer

GDS uses DMA transfers between GPU memory and storage, avoiding a bounce buffer through the CPU.

Answer

GDS always routes storage data through a CPU bounce buffer to validate it before the GPU can read.

Answer

NVIDIA recommends the legacy nvidia-peermem module over DMA-BUF for all GDS deployments here.

Question 4

An engineer is designing a Triton Inference Server deployment and must confirm how client inference requests reach the correct model. Which statement matches the Triton architecture?

Accepted Answer

Requests arrive via HTTP/REST, gRPC, or the C API and are routed to the per-model scheduler.

Answer

Requests must use gRPC only and are broadcast to every loaded model before backend selection.

Answer

Requests are written to the model repository on disk and polled by each backend in a loop.

Answer

Requests bypass any scheduler and are sent straight to a single shared global batching queue.

Question 5

A platform team is evaluating cuML to speed up an existing CPU machine learning workflow. Which TWO statements about cuML are documented? (Select two.)

Accepted Answer

With cuml.accel it can accelerate existing code on the GPU with zero code changes needed

Accepted Answer

It delivers on average 10-50x faster performance than CPU-based alternatives on workloads

Answer

It requires rewriting models into low-level CUDA kernels before any real speedup is realized

Answer

It is limited only to clustering and cannot perform regression or classification tasks

Question 6

An MLOps team is investigating Triton Inference Server's Pending Request Count to size model concurrency: they send 100 requests to a model instance that can only execute 5 concurrently. According to the Triton metrics documentation, what should the Pending Request Count show?

Accepted Answer

It should show about 95, because it counts requests received but not yet started execution by a backend model instance.

Answer

It should show 100, because every request received via TRITONSERVER_InferAsync is counted as pending until the entire batch fully completes.

Answer

It should show 5, because the metric reports only the requests currently executing on the backend model instance at that moment.

Answer

It should show 0, because Pending Request Count tracks failed requests rather than the ones still waiting in the per-model queue.

Question 7

An organization needs a universal NVIDIA system that handles analytics, training, and inference on the same chassis. Which DGX statement matches the introduction?

Accepted Answer

The NVIDIA DGX H100/H200 Systems are the universal systems purpose-built for all AI infrastructure and workloads from analytics to training to inference.

Answer

DGX H100/H200 systems are intentionally limited to inference-only workloads and cannot be used for training jobs.

Answer

DGX H100/H200 systems are CPU-only platforms and rely on external GPU appliances for any AI work.

Answer

DGX H100/H200 systems are sold without GPUs; customers supply their own H100 cards from third parties.

Question 8

A computer vision team is hitting a wall where GPU utilization stalls because JPEG decode and augmentation run on the CPU. Which NVIDIA library is documented to move that data loading and pre-processing onto the GPU?

Accepted Answer

DALI, a GPU-accelerated library for data loading and pre-processing of image, video, and audio

Answer

cuDNN, a GPU-accelerated library of primitives such as convolution, attention, and pooling layers

Answer

cuVS, a GPU-accelerated library for approximate and exact nearest-neighbor vector similarity search

Answer

NGC, a hub of GPU-optimized containers and pretrained models for data science and HPC teams

Question 9

An engineer is evaluating whether the older Volta-generation V100 GPUs in a cluster can be partitioned with Multi-Instance GPU before standardizing on MIG. According to the GPU Operator MIG documentation, which GPU architectures support MIG partitioning?

Accepted Answer

MIG is supported on NVIDIA Ampere and later architectures such as the A100, so the Volta V100 GPUs cannot use MIG.

Answer

MIG is supported on every CUDA-capable NVIDIA GPU, so the Volta V100 GPUs can be partitioned just like an A100 can.

Answer

MIG is supported only on the Volta architecture, so the V100 GPUs are actually the ideal hardware for MIG partitioning.

Answer

MIG is supported on Kepler and all newer datacenter GPUs, so the V100 GPUs qualify once the R450 driver is installed.

Question 10

A developer is being onboarded to NCCL and has a fixed mapping of ranks to CUDA devices for a single-process, multi-GPU job. Which functions does the documentation list for creating the communicator objects?

Accepted Answer

ncclCommInitRank(), ncclCommInitRankConfig(), and ncclCommInitAll() create them

Answer

ncclGroupStart() and ncclGroupEnd(), which themselves build the communicators

Answer

ncclCommSplit() and ncclCommFinalize(), the only fixed-rank constructors

Answer

ncclBroadcast() and ncclAllReduce(), creating the communicator on first call

Question 11

A production team is choosing inference middleware for a fleet that mixes TensorRT, PyTorch, and ONNX models. Which statement best captures the Triton Inference Server framework coverage?

Accepted Answer

Triton enables teams to deploy any AI model from multiple deep learning and machine learning frameworks, including TensorRT, PyTorch, ONNX, OpenVINO, Python, RAPIDS FIL, and more.

Answer

Triton must be recompiled per framework with framework-specific binaries and cannot host a mixed model repository at runtime.

Answer

Triton only serves TensorRT engines and refuses any other backend at server start-up time.

Answer

Triton accepts PyTorch models only when they have been converted into a proprietary NVIDIA TensorFormat that is incompatible with ONNX exports.

Question 12

An architect is deciding between GPU time-slicing and Multi-Instance GPU (MIG) for a multi-tenant cluster where each tenant must be protected from another tenant's memory faults and overuse. Which statement correctly captures the trade-off documented for time-slicing?

Accepted Answer

Time-slicing provides no memory or fault isolation between replicas, so MIG is the appropriate choice when tenants require hardware isolation.

Answer

Time-slicing gives each replica its own fault-isolated memory partition, so it satisfies the tenant-isolation requirement without needing MIG.

Answer

Time-slicing and MIG provide identical memory and fault isolation, so the choice here is purely about which Kubernetes scheduler is in use.

Answer

Time-slicing isolates compute but shares memory, so MIG is only required when the tenants additionally need separate CUDA driver versions.

Question 13

A platform team needs to confirm what the H100 L2 cache size is relative to A100 before sizing kernels for SuperPOD. Which value matches the Hopper documentation?

Accepted Answer

The L2 cache provides 50 MB capacity, up from A100's 40 MB.

Answer

The L2 cache is 6 MB on H100, the same as Volta's V100 L2 cache.

Answer

H100 eliminated the L2 cache entirely; only the L1 cache remains.

Answer

The L2 cache on H100 shrank to 20 MB to make room for additional Tensor Cores.

Question 14

An organization needs to estimate whether Datashader can handle far larger point datasets than fit on one laptop and asks about its documented scaling behavior. Which statement matches Datashader's documentation?

Accepted Answer

It can plot about a billion points in a second on a 16GB laptop and scales up to out-of-core, distributed, or GPU processing for bigger data.

Answer

It is hard-capped at only a few thousand points per chart and cannot move beyond the memory of a single laptop in any configuration.

Answer

It runs only on multi-node CPU clusters and cannot use a single laptop nor any GPU acceleration to render large point datasets at all.

Answer

It offloads all rendering to a hosted cloud API and is unable to rasterize point data locally on the machine that holds the data.

Question 15

A platform team is evaluating GPU time-slicing through the NVIDIA GPU Operator to let several low-priority training jobs share oversubscribed GPUs on a Kubernetes cluster. According to the GPU Operator documentation, what does time-slicing actually do for those workloads?

Accepted Answer

It schedules the oversubscribed GPU's replicas so workloads from different pods interleave their execution on the same underlying GPU.

Answer

It physically partitions each GPU into fixed hardware slices so every job receives a guaranteed, isolated share of memory and compute.

Answer

It migrates pending jobs to idle GPUs on other nodes so that no single GPU is ever oversubscribed during the cluster's peak demand.

Answer

It throttles each pod's clock speed proportionally so that the total GPU power draw stays under a fixed cluster-wide energy budget.

NVIDIA Certified Associate - AI Infrastructure and Operations

Sample questions

Sources

Similar exams