NVIDIA-Certified Associate: AI Infrastructure and Operations
Independent: not affiliated with NVIDIA. Free to try; the study bundle adds a study guide and two practice exams.
5 original questions, none of them in the practice exams.
Pick an answer. Show answer explains every option and links its NVIDIA source.
5 questions / Explanations link NVIDIA sources / Last updated Oct 26, 2026
Question 1
Multiple response
Domain 1 / Task 1.2 Training versus inference requirements
A payments company has trained a fraud-detection model and will now run it in production around the clock, scoring each card transaction while the customer waits for approval. The infrastructure team is choosing storage for the production system. Which two requirements should drive that choice? Choose two.
Show answerHide answer
Answers: B and E
Why
Inference is where the value of an AI solution is delivered, and NVIDIA's storage guidance treats its needs separately from training's. For real-time fraud detection it calls for submillisecond storage and a high-performance path between storage and GPU memory, naming GPUDirect Storage, and it warns that apps such as fraud detection have no good time for a maintenance window, so the storage should support upgrades without downtime.
Why not the others
ANVIDIA's guidance for inference is scale-out storage that grows performance along with capacity; a scale-up design adds capacity without adding performance.
CA running inference service also writes data, such as customer profiles and other records, plus data kept for retraining, so its storage has to grow in capacity and performance.
DStreaming the full dataset into GPU memory pass after pass is a training need; this system scores live transactions.
A systems administrator knows that on a CPU server, running many more threads than there are cores slows work down, because the operating system keeps swapping threads on and off the cores. A GPU job on the same server keeps tens of thousands of threads active on one GPU without that penalty. What explains the difference?
Show answerHide answer
Answer: B
Why
NVIDIA's CUDA Best Practices Guide contrasts the two designs. CPU threads are heavyweight: the operating system swaps them on and off the cores, so context switches are slow and expensive. GPU threads are lightweight: thousands are queued in warps of 32, and when one warp has to wait, the GPU starts work on another. Because every active thread has its own registers, nothing is swapped. CPU cores are built to minimize latency for a few threads at a time, GPUs to maximize throughput across many.
Why not the others
AGPU threads do wait on memory; the GPU covers the wait by running other warps. NVIDIA contrasts this latency-hiding design with cache-based CPUs.
CThe GPU's advantage is how many threads it keeps in flight, not the speed of any one thread; CPU cores are the ones designed to minimize latency per thread, and GPU threads are switched with no swap at all.
DNo operating system swaps GPU threads: the GPU moves between warps itself, and resources stay allocated to each thread until it completes.
Domain 2 / Task 2.5 Components of an accelerated cluster
A DGX H100 system arrives with its two OS drives in a RAID 1 array and its eight 3.84 TB NVMe data drives in a RAID 0 array that serves as the system's local data cache. The cluster's training datasets are kept on a shared parallel file system. A storage administrator wants to rebuild the eight data drives as mirrored RAID 1 pairs so that a failed drive cannot lose data. What should the infrastructure lead advise?
Show answerHide answer
Answer: C
Why
NVIDIA's DGX H100 user guide lists the OS storage as two NVMe M.2 SSDs in a RAID 1 array and the data cache as eight 3.84 TB NVMe U.2 drives in a RAID 0 array. The DGX OS user guide says the data drives 'can be configured as RAID-0 for performance or RAID-5 for resiliency', that 'RAID-0 is recommended for data caching' and that 'RAID-5 should be used for persistent data storage'. The cache speeds up re-reads of data whose primary copy lives on the cluster's shared storage; if a cache drive fails, the array is re-created and the data is read again from that storage, so nothing is lost. Striping across all eight drives gives the cache their full capacity and combined throughput, while mirrored pairs would halve the usable capacity to protect copies that are not at risk.
Why not the others
AThe primary copy of the training data is on the shared parallel file system; the local array holds cached copies that can be read again after a failure.
BNVIDIA's DGX OS user guide says 'The RAID level for the data drive can be changed after the installation of Ubuntu', although the change destroys the data on the drives. RAID 0 stays because it is the level NVIDIA recommends for a cache, not because it is locked.
DThe local cache is there to speed up re-reads; without it, every epoch reads the data again over the storage network from the shared file system.
A training job runs on all eight GPUs of one DGX H100 system and spends much of each step exchanging gradients among them. An engineer suggests that faster cluster network cards, 800 Gb/s InfiniBand instead of the system's 400 Gb/s ConnectX-7 cards, would speed the job up. Why will that not help this job?
Show answerHide answer
Answer: D
Why
The DGX H100 user guide lists NVSwitch with fourth-generation NVLink giving 900 GB/s of GPU-to-GPU bandwidth inside the system, and eight single-port ConnectX-7 InfiniBand cards of up to 400 Gb/s each as the cluster network. Gradient traffic among one system's GPUs uses the NVLink path, so the cluster cards, and their speed, are not on this job's path; they matter once a job spans several systems.
Why not the others
ASHARP in InfiniBand switches speeds up collectives between systems; it does not move traffic among one system's GPUs onto the cards.
BThe guide lists separate ConnectX-7 cards for storage and in-band management; the InfiniBand cards are the cluster network, which carries training traffic between systems.
CThe gradients move between the GPUs over NVLink inside the system; adding CPU cores does not change that path.
Domain 3 / Task 3.4 Considerations for virtualizing accelerated infrastructure
A hospital runs a GPU-accelerated imaging VM on each of eight hypervisor hosts, and each VM has a whole GPU assigned through passthrough. The virtualization team must apply host firmware updates that need a reboot, and before each reboot it has to move that host's running VMs to other hosts without shutting them down. What change makes this possible?
Show answerHide answer
Answer: A
Why
NVIDIA's vGPU FAQ compares the two. With passthrough, an entire GPU is assigned to one VM, which has exclusive access; that gives maximum performance but does not support live migration or suspend/resume. vGPU mode supports both, so on a hypervisor that supports vGPU migration the VMs can be moved off a host before it reboots.
Why not the others
BNVIDIA states that passthrough does not support live migration: the whole GPU belongs to that one VM.
CNVIDIA states that passthrough does not support suspend/resume either.
DThe firmware update reboots the whole host, so a second GPU in the same host does not keep the VM running, and passthrough still cannot move it.
The study bundle has two practice exams with 100 original questions, an answer key and an explanation for every answer, plus the 56-page study guide. Practice exam explanations are shorter than the ones on this page: a short paragraph per question, with no source links.
Study bundle
$19
100 practice questions / Study guide / Answer keys
Every bundle includes 6 months of access: download the PDFs, and any new editions, until 6 months after your purchase date. Your receipt has a license key. Enter it at overqual.com/download for the three PDFs, stamped with your email and order number. They are yours to keep.