The NVIDIA H100 GPU supports up to 7 Multi-Instance GPU (MIG) partitions, allowing it to be divided into seven isolated instances for multi-tenant or mixed workloads. This capability leverages the H100’s architecture to maximize resource flexibility and efficiency, with 7 being the documented maximum.

(Reference: NVIDIA H100 GPU Documentation, MIG Section)

The correct answer is B because distributed AI training performance depends heavily on high-bandwidth, low-latency inter-node communication. NVIDIA DGX SuperPOD reference architecture states that InfiniBand “continues to evolve and lead data center network performance,” with NDR InfiniBand providing “400 Gbps per direction” and “extremely low port-to-port latency.” It also notes that InfiniBand provides additional performance-optimization features, including adaptive routing and collective communication with NVIDIA SHARP.

NVIDIA Network Operator documentation also states that it delivers “high-throughput, low-latency networking for scale-out, GPU computing clusters” and that RDMA supports memory-to-memory transfers that “bypass the CPU and kernel networking stack,” with support for InfiniBand and RoCE protocols. This directly supports deploying a lossless InfiniBand or RoCE fabric for distributed training traffic such as all-reduce communication.

Why the other options are incorrect: Wi-Fi is unsuitable for high-performance multi-rack GPU training communication. Stateful firewalls and deep-packet inspection between training nodes would add latency and bottlenecks. Adding switch ports without fixing oversubscription and latency does not solve distributed all-reduce scaling inefficiency.

GPUDirect RDMA is used in multi-server GPU workloads to enable a direct peer-to-peer data path between GPU memory and NVIDIA networking devices. NVIDIA states that GPUDirect RDMA provides “a direct P2P data path” between GPU memory and NVIDIA host networking devices, which reduces GPU-to-GPU communication latency and “completely offloads the CPU.”

This means the direct benefit is that CPU involvement in GPU-to-GPU network communication is removed or greatly reduced. The option “Offloads data movement from CPUs” is therefore correct. NVIDIA’s GPUDirect page also explains that network adapters and storage drives can directly read and write GPU memory, “eliminating unnecessary memory copies,” decreasing CPU overhead, and reducing latency.

Why the other options are incorrect: GPUDirect RDMA does not raise GPU memory clock speeds, does not primarily act as a CPU scheduling feature, and does not compress transferred data. Its purpose is direct data movement between GPU memory and network/storage devices to reduce latency, reduce unnecessary copies, and lower CPU overhead.

RoCE is the correct answer because it provides RDMA over Ethernet for low-latency, efficient data movement. NVIDIA networking documentation states: “Remote Direct Memory Access (RDMA) is the remote memory management capability that allows server-to-server data movement directly between application memory without any CPU involvement.” It then states: “RDMA over Converged Ethernet (RoCE) is a mechanism to provide this efficient data transfer with very low latencies on lossless Ethernet networks.” NVIDIA DOCA documentation similarly states that RoCE extends RDMA functionality to lossless Ethernet networks, delivering “high-throughput, ultra-low latency communication.”

This is especially important for large AI clusters because distributed training requires fast GPU-to-GPU and node-to-node communication. NVIDIA states that Spectrum-X builds on Ethernet with RoCE extensions to enhance performance for AI, bringing InfiniBand-style best practices such as adaptive routing and congestion control to Ethernet.

Why the other options are incorrect: DCTCP and ECN can support congestion control, but they are not the core GPU-to-GPU low-latency data-transfer protocol. PFC-only Ethernet without RDMA does not provide the direct memory-access benefit. iWARP is RDMA over TCP, but NVIDIA AI Ethernet designs emphasize RoCE for high-performance AI networking.

Training architecture generally requires more data storage because training uses large datasets, repeated data access, and checkpointing. NVIDIA DGX SuperPOD storage architecture documentation explains that storage performance requirements vary by model and dataset, and it specifically highlights training workloads, datasets, and checkpoint files as important storage design considerations. For large model use cases, NVIDIA notes that “peak performance for reads and writes are needed for creating and reading checkpoint files” and that training stops during checkpoint operations. NVIDIA also states that storage solutions for AI factories must support “large-scale AI training, fine-tuning, inference, KV cache, and retrieval-augmented generation,” showing that storage is relevant across the lifecycle but especially demanding for training and fine-tuning.

Inference architecture typically stores the deployed model, runtime components, and sometimes cache or retrieval data. Training architecture must store training datasets, intermediate outputs, logs, and checkpoints, so it generally has the greater storage requirement.

Multi-node training is advantageous when a model’s size—its parameters, activations, and gradients—exceeds the memory capacity of a single GPU. By sharding the model across multiple nodes (using techniques like data parallelism or model parallelism), training becomes feasible and efficient. User count and inference scale are unrelated to training architecture needs, which focus on compute and memory distribution.

(Reference: NVIDIA AI Infrastructure and Operations Study Guide, Section on Multi-Node Training Benefits)

Kubernetes excels in container orchestration with advanced scheduling (assigning workloads based on resource needs and availability) and load balancing (distributing traffic across pods via Services). It’s not inherently bare-metal (it runs on various platforms), and inferencing capability depends on applications, not Kubernetes itself, making B and D the true statements.

(Reference: NVIDIA AI Infrastructure and Operations Study Guide, Section on Kubernetes Orchestration)

Training neural networks typically involves streaming large datasets from storage in a sequential read pattern. This ordered access maximizes throughput and minimizes seek overhead, as training pipelines ingest data in batches for processing across epochs. Writes (e.g., model checkpoints) are less frequent and typically sequential, while random writes are rare, making sequential reads the dominant pattern. (Note: The document incorrectly lists C as the answer; B aligns with NVIDIA’s documentation.)

(Reference: NVIDIA AI Infrastructure and Operations Study Guide, Section on Storage Access Patterns)

Cloud GPU infrastructure lowers the cost barrier to entry by offering a pay-as-you-go model, eliminating the need for significant upfront capital expenditure on hardware. While on-prem may offer I/O cost savings or hardware control, the cloud’s accessibility and reduced initial investment make it a compelling choice for organizations seeking immediate GPU access without large sunk costs.

(Reference: NVIDIA AI Infrastructure and Operations Study Guide, Section on Cloud GPU Advantages)

The nvidia-smi (System Management Interface) command is the primary tool for checking GPU utilization on NVIDIA systems, including the DGX H100. It provides real-time metrics like utilization percentage, memory usage, and power draw. NVML (NVIDIA Management Library) is an API, not a command, and ctop is unrelated, solidifying nvidia-smi as the standard.

(Reference: NVIDIA DGX H100 System Documentation, Monitoring Section)