Free NCP-AIO practice test: 30 questions with answers

A managed configuration or GPU operator keeps driver and runtime versions consistent across nodes. Manual, user-level, or disabled GPU approaches reintroduce inconsistency or remove capacity.

InfiniBand and RDMA-capable Ethernet provide the bandwidth and low latency needed for distributed training traffic. Dial-up, serial cables, and older Wi-Fi do not meet that requirement.

MIG splits supported GPUs into isolated instances with dedicated memory and compute. Tensor Cores are compute units, NVLink is a GPU interconnect, and NVENC is a video encoder.

The GPU Operator automates driver, container toolkit, and device plugin deployment for GPU nodes in Kubernetes. The other products are end-user applications.

Confirming the driver and that the device plugin reports GPUs as allocatable identifies whether the GPUs are available to the scheduler. Chassis replacement, hostnames, and disk size are unrelated.

DCGM collects GPU telemetry and exporters make it available to monitoring systems. Calculators, text editors, and printer utilities do not monitor GPUs.

Idle GPUs with queued jobs indicate scheduling or input bottlenecks such as slow data loading. A fully utilized cluster would not show idle GPUs while work waits, and printers and cable lengths are unrelated.

DCGM exporters with a metrics pipeline and dashboards give continuous, cluster-wide visibility. Printed sheets, monthly manual checks, and user reports are late or incomplete.

Interconnect utilization and time spent communicating per step show whether communication limits training. Display brightness, printed pages, and logins are unrelated.

Average utilization and queue wait time show how efficiently GPUs are used and whether demand exceeds capacity. Printers, tile colours, and email accounts are unrelated.

Steadily growing memory usage often indicates a leak or cache growth that could eventually cause out-of-memory failures. The other conclusions are not supported by the pattern.

Alerts on GPU errors or missing GPUs catch hardware faults early. Login, lighting, and file-rename events do not indicate GPU health.

Schedulers with queues, quotas, and priorities provide fair sharing and efficient use. Unlimited first-come allocation, email-based allocation, and forcing all jobs onto one GPU are unfair or inefficient.

A job that needs four GPUs should wait until the request can be satisfied, which keeps results correct. Silent partial runs, cancellation without notice, and misreporting CPU runs are incorrect.

Distributed training needs all workers to start together, so gang scheduling avoids partial starts that block or waste GPUs. It does not reduce GPU count, networking needs, or model size.

Time limits and preemption keep resources flowing to others and enforce policy. Unlimited runs, disabled schedulers, and maximum allocations make blocking worse.

Workload size, project priority, and fair share determine a sensible quota. Logo colours and fonts are unrelated to resource allocation.

A GPU resource request and a volume mount give the pod exactly the GPUs and data it needs. Printers, labels alone, and DNS entries do not allocate GPUs or mount data.

Private networking and locally hosted registries with access controls keep data and dependencies inside the boundary. Unrestricted internet access, public buckets, and disabled security expose the data.

An allowlist, admission policy, trusted registries, and scanning limit what can run. Allowing any image, trusting names alone, and root containers each increase risk.

Namespaces with quotas, RBAC, and network policies isolate tenants on shared infrastructure. Shared admin accounts, lack of isolation, and open data directories expose tenants to one another.

RBAC, audit logs, and access reviews support governance and accountability. Shared passwords and disabled monitoring weaken control.

Audit logs record administrative changes with identity and time, which supports accountability. Printed binders, verbal accounts, and reminders do not provide reliable evidence.

Encryption at rest, role-based access, and logging protect sensitive models. Plain-text shares, public links, and personal email expose them.

Communication timeouts usually point to network health or connectivity on the training fabric, so that is checked first. Log colour, stickers, and author names do not affect communication.

Draining workloads and diagnosing before replacing the GPU protects jobs and confirms the fault. Continuing on a faulty GPU, ignoring errors, and disabling logs risk corrupted results.

Differences in clocks, temperature, or drivers commonly explain node-to-node performance gaps. Wallpapers, hostnames, and printers do not affect GPU performance.

Link speed, MTU mismatches, and error counters are common causes of throughput loss after a switch change. Restarting GPUs, reinstalling systems, and changing architecture do not address the network change.

A GPU with less free memory, due to other processes or a leak, can fail where others succeed. Speed, name length, and layer count do not explain memory failures on one GPU.

After a reboot, the driver must load and the runtime must be configured to expose GPUs to containers. Printer drivers, hostnames, and image deletion do not restore GPU access.