NCP-AIO Installation and Deployment — 30 Questions

Question 1

Which TWO steps are required to prepare a node for NVIDIA vGPU deployment on an ESXi cluster?

Accepted Answer

Set the GPU graphics device mode to 'Shared' in the host vSphere settings.. Preparing an ESXi host for vGPU requires both hardware-level configuration and software-level driver deployment. Enabling the Graphics Device in shared mode is the prerequisite for the hypervisor to present the GPU to virtual machines, while installing the NVIDIA VIB (vSphere Installation Bundle) provides the necessary communication layer between the guest OS and the physical hardware. Correct orchestration here ensures that the hypervisor can successfully broker GPU resources.

Answer

Set the GPU graphics device mode to 'Shared' in the host BIOS.

Answer

Upgrade the guest VM virtual hardware to version 15 or lower.

Answer

Disable the NVIDIA License System on the host.

Question 2

Which THREE items must be verified before installing the NVIDIA GPU Operator on a Kubernetes cluster?

Accepted Answer

The worker nodes must have sufficient GPU resources and capacity.. The GPU Operator relies on specific cluster conditions to function. Verify the kernel version to ensure it is supported by the driver, ensure nodes have correct labels for scheduling, and confirm that the Kubernetes API version meets the minimum threshold for the Operator's Custom Resource Definitions. These prerequisites are vital for successful deployment, as they prevent common failures during the lifecycle management of GPU resources within complex containerized orchestration environments.

Answer

NVIDIA drivers must be pre-installed on the host OS.

Answer

The Kubernetes cluster must use the Helm 2 package manager.

Question 3

Which TWO of the following are necessary to correctly monitor GPU health using the NVIDIA GPU Operator?

Accepted Answer

Deploy the DCGM Exporter via the GPU Operator.. Monitoring requires both a data collector and a visualization engine. The NVIDIA DCGM Exporter collects the metrics from the GPU, and a monitoring stack like Prometheus (often used alongside Grafana) provides the storage and graphing capabilities. This combination is essential for AI Ops to detect performance degradation or hardware failures before they impact production training models, allowing for proactive maintenance and better resource utilization tracking.

Answer

Install a local SNMP daemon on each GPU node.

Answer

Configure the host kernel to support NVIDIA-SMI logging.

Answer

Enable the Kubernetes Dashboard for all nodes.

Question 4

An administrator is planning an NVIDIA Base Command Manager deployment. Which TWO factors are critical to consider for successful node provisioning?

Accepted Answer

The BIOS and BMC firmware versions must be identical across all nodes to ensure consistent provisioning behavior.. Successful provisioning in DGX environments relies on accurate network topology and firmware consistency. These factors ensure that the management software can correctly orchestrate workloads across the cluster. Understanding these requirements is vital for AI operations because failed provisioning results in stranded compute resources, preventing the scaling of large models that require multi-node synchronization and high-bandwidth interconnectivity across the entire GPU cluster infrastructure.

Answer

The deployment network should be configured as a flat Layer 2 network without VLAN tagging.

Answer

The NVIDIA drivers should be installed manually on each node after the OS is provisioned via the network.

Answer

GPU memory settings must be configured via the BIOS for each node manually before provisioning starts.

Question 5

Which THREE steps are necessary for properly configuring GPUDirect RDMA in a multi-node AI cluster deployment?

Accepted Answer

Ensure the PCIe slot supports peer-to-peer communication between the NIC and the GPU.. GPUDirect RDMA is vital for bypassing CPU bottlenecks during high-bandwidth communication between GPUs. These configuration steps ensure that networking hardware can directly access GPU memory buffers. For AI professionals, this is essential to reduce latency and improve scaling efficiency, as it minimizes data copying between host memory and GPU memory, which is the primary bottleneck in distributed training jobs.

Answer

Enable the NVIDIA Unified Memory setting in the system BIOS.

Answer

Set the GPU persistence mode to off to allow the driver to manage memory more dynamically.

Question 6

During a firmware update on a DGX H100 system, the BMC becomes unresponsive. What is the standard recovery procedure an AI operations professional should follow?

Accepted Answer

Initiate a hard power cycle by disconnecting all power supplies for at least 60 seconds.. Managing BMC failures requires a controlled power cycle to force a controller reset, independent of the host OS. This is a critical skill because BMCs provide out-of-band management necessary for remote troubleshooting. If the BMC remains unresponsive, the hardware cannot be monitored or managed, leading to potential downtime and inability to perform further updates or diagnostics during critical training windows.

Answer

Run the firmware update utility again from the host OS to overwrite the corrupted BMC state.

Answer

Use the UEFI shell to reflash the BMC firmware directly using a network mount.

Answer

Reinstall the host operating system to clear the BMC management drivers.

Question 7

An administrator needs to ensure that all GPUs in a cluster have the latest ECC error reporting enabled. Which command should they execute to check the current ECC status?

Accepted Answer

nvidia-smi -q -d ECC. Checking ECC status is vital for data integrity in AI training. ECC (Error Correction Code) helps detect and recover from bit-flips, preventing corrupted model weights during long training runs. Regular monitoring of ECC logs ensures that hardware degradation is identified early, preventing silent failures that could invalidate weeks of expensive compute and model training time.

Answer

nvidia-smi --list-gpus

Answer

nvidia-smi --ecc-config=on

Answer

nvidia-smi --help-all

Question 8

Which TWO actions should be performed when replacing a failed GPU in a DGX system?

Accepted Answer

Verify that the GPU is detected by the OS using the lspci command.. GPU replacement requires both physical care and software-level verification to ensure the new hardware is integrated into the cluster. Failing to properly re-initialize the driver or verify the hardware state can lead to performance imbalances or system instability. Following these steps ensures that the new component is fully recognized and operating correctly, allowing for seamless resumption of high-compute AI workloads.

Answer

Update the BMC firmware immediately after physical installation to match the other GPUs.

Answer

Manually format the GPU local memory to remove any remaining data from the previous card.

Answer

Change the GPU BIOS settings to allow higher voltage for the new card.

Question 9

Refer to the exhibit. An AI engineer is reviewing a compliance policy for a new deployment. Based on the JSON policy provided, which condition must be met for a node to be considered compliant?

Accepted Answer

The driver version must be 535.54.03 or higher, and the CUDA version must be 12.2.. Compliance policies are essential for maintaining a standardized software stack across a cluster. This specific policy mandates a minimum driver version and a specific CUDA version. Adherence to these versions ensures that all nodes have consistent capabilities and behaviors, which is critical for large-scale distributed training where version discrepancies can lead to non-deterministic results or library incompatibilities.

Answer

The driver version must be exactly 535.54.03.

Answer

The CUDA version can be any version as long as the driver is updated.

Answer

The driver version must be lower than 535.54.03 to ensure backward compatibility.

Question 10

When deploying NVIDIA Base Command Platform, which networking protocol is primarily used for node-to-node management traffic?

Accepted Answer

IPMI. IPMI (Intelligent Platform Management Interface) is the standard for out-of-band management of servers. It allows administrators to perform power-on, reset, and monitoring tasks even if the OS is not running. In the context of NVIDIA Base Command, IPMI is fundamental for the platform to manage and orchestrate hardware resources across the cluster, ensuring that nodes can be remotely recovered or configured as needed.

Answer

InfiniBand

Answer

NVLink

Answer

RoCE

Question 11

An administrator is optimizing a cluster for NCCL operations. Which kernel setting is often recommended to improve performance by increasing the number of pinned memory regions?

Accepted Answer

ulimit -l (memlock). NCCL uses shared memory and pinned memory regions to achieve high-speed communication between GPUs. Increasing the limit for locked memory (mlock) allows the system to pin more memory, which reduces CPU overhead during data transfers. For AI operations, this is a standard tuning step to prevent bottlenecks in large-scale distributed training where GPU-to-GPU communication is the primary performance factor.

Answer

vm.max_map_count

Answer

fs.file-max

Answer

net.core.rmem_max

Question 12

An administrator is planning an NVIDIA AI Enterprise deployment on VMware vSphere. Which TWO requirements must be met to support Multi-Instance GPU (MIG)?

Accepted Answer

NVIDIA Ampere or later architecture GPUs. MIG partitioning requires specific hardware support and virtualization configurations. Ensuring the underlying GPU architecture supports hardware-level partitioning is the primary constraint. Furthermore, the hypervisor must be configured to pass through the hardware correctly, allowing the NVIDIA Virtual GPU manager to slice the physical GPU into isolated instances. These requirements ensure that secure, performance-guaranteed environments are available for diverse AI workloads.

Answer

NVIDIA vGPU license on every partition

Answer

Standard NVIDIA vGPU driver in the guest

Answer

Kubernetes ingress controller

Question 13

Which tool provides a centralized dashboard for monitoring GPU utilization, health, and workload performance within an NVIDIA AI Enterprise environment?

Accepted Answer

NVIDIA DCGM Exporter. NVIDIA DCGM Exporter, when paired with Prometheus and Grafana, serves as the standard toolset for monitoring GPU metrics. It collects data from the Data Center GPU Manager (DCGM) and exposes it for visualization. This is critical for AI Ops as it allows administrators to detect performance bottlenecks, monitor thermals, and track GPU memory usage across the entire cluster lifecycle.

Answer

NVIDIA NGC CLI

Answer

NVIDIA CUDA Profiler

Answer

NVIDIA vGPU Software Manager

Question 14

An admin is deploying AI Enterprise on OpenShift. Which THREE tasks must be completed to ensure the GPU Operator functions correctly?

Accepted Answer

Create a MachineConfigPool for GPU nodes. Deploying the GPU Operator on OpenShift requires tight integration with the platform's security and node-configuration features. The operator needs permissions to modify node configurations and must have the correct machine config pools created to handle driver installations. These steps ensure that the operator can effectively manage the lifecycle of drivers and plugins within the strict security boundaries defined by the OpenShift platform.

Answer

Manually install CUDA on all nodes

Answer

Configure a local image mirror

Question 15

An administrator wants to ensure that specific pods are prioritized for GPU access during high-demand training sessions. Which Kubernetes feature should they configure?

Accepted Answer

Node Affinity. Kubernetes Taints, Tolerations, and Node Affinity allow administrators to control which pods land on which nodes. By dedicating specific GPU nodes to high-priority workloads and applying taints, admins prevent non-critical pods from consuming GPU capacity. This ensures that high-priority training jobs have immediate access to the necessary hardware, optimizing throughput and meeting performance SLAs in production AI environments.

Answer

Horizontal Pod Autoscaler

Answer

ClusterIP Service

Answer

Persistent Volume Claims

Question 16

Which TWO actions should be performed to optimize GPU performance for a multi-tenant AI workload?

Accepted Answer

Enable MIG (Multi-Instance GPU). To optimize performance in a multi-tenant environment, isolation and scheduling efficiency are key. MIG allows for hardware-level isolation, preventing one tenant from impacting another, while proper node scheduling ensures that workloads are placed on the most appropriate hardware. Together, these strategies maximize hardware efficiency and protect the performance of critical workloads from noisy neighbors in a shared cluster environment.

Answer

Reduce the Kubernetes API polling interval

Answer

Disable the NVIDIA Container Toolkit

Answer

Use the standard Linux kernel scheduler

Question 17

An administrator is deploying NVIDIA AI Enterprise on a bare-metal server. Which component is strictly required to enable GPU virtualization for multiple virtual machine instances?

Accepted Answer

NVIDIA vGPU Manager. NVIDIA AI Enterprise relies on the NVIDIA vGPU software stack to slice a single physical GPU into multiple virtual instances. Installing the vGPU manager on the hypervisor is the foundational step for orchestration, allowing VMs to access GPU resources via the guest driver. Without this virtualization layer, instances would lack the necessary hardware acceleration APIs required for intensive AI inference or training workloads in multi-tenant environments.

Answer

NVIDIA Base Command

Answer

NVIDIA Collective Communications Library (NCCL)

Answer

CUDA Toolkit

Question 18

Which environment variable must be set in a container's configuration to ensure the application uses a specific GPU when multiple are available?

Accepted Answer

NVIDIA_VISIBLE_DEVICES. The CUDA_VISIBLE_DEVICES environment variable is the standard mechanism for controlling GPU visibility in applications. By setting this variable, an administrator can restrict a container process to only see and use specific physical GPUs. This is fundamental for managing resource assignment in multi-GPU nodes, ensuring that workload scheduling is enforced and that processes do not interfere with each other by grabbing unassigned devices.

Answer

CUDA_DEVICE_ORDER

Answer

GPU_DEVICE_INDEX

Answer

KUBERNETES_GPU_LIMIT

Question 19

Which NVIDIA container runtime configuration is mandatory for ensuring that containers can successfully interact with physical GPU hardware within a standard Docker or Kubernetes environment?

Accepted Answer

nvidia-container-runtime. The NVIDIA container runtime acts as a bridge between the containerized application and the host's NVIDIA driver. It intercepts the container creation process to inject the necessary GPU device nodes and libraries (like CUDA) into the container's environment. Without this runtime, the container lacks the low-level visibility required to interface with the GPU, leading to runtime errors when the application attempts to execute GPU-accelerated code or libraries.

Answer

runc

Answer

crun

Answer

kata-runtime

Question 20

An administrator is deploying NVIDIA AI Enterprise on a bare-metal server. Which prerequisite must be verified to ensure the NVIDIA AI Enterprise suite operates correctly with the underlying hardware acceleration?

Accepted Answer

The host kernel version must be verified against the NVIDIA Driver compatibility matrix.. Verifying that the installed GPU driver version matches the specific NVIDIA AI Enterprise release requirements is critical. Mismatched drivers lead to kernel module failures or inability to access GPU compute resources, preventing container orchestration platforms like Kubernetes from scheduling AI workloads effectively. This verification step ensures the software stack correctly interfaces with the NVIDIA GPU hardware, maintaining the performance and stability required for enterprise-grade AI production environments.

Answer

The server BIOS must support UEFI Secure Boot for all GPU firmware updates.

Answer

The server must have an active connection to the NVIDIA NGC private registry.

Answer

The GPU must be partitioned into at least four Multi-Instance GPU (MIG) slices.