Server and Network Installation and Configuration
Free NVIDIA-Certified Professional: AI Infrastructure practice — 6 questions on Server and Network Installation and Configuration, with explanations. No sign-up. Full 12-question mixed test →
A DGX H100 SuperPOD scalable unit uses a rail-optimized InfiniBand fat-tree. Each DGX H100 node has 8 ConnectX-7 adapters dedicated to the compute fabric. To preserve rail optimization, how should these 8 adapters be cabled to the leaf switch layer?
Rail-optimized design dedicates one leaf switch per adapter rank (rail) across all nodes in the SU. Adapter 0 on every DGX connects to leaf switch 0, adapter 1 to leaf switch 1, and so on. This keeps GPU-to-GPU traffic for a given rail on a single switch, minimizing hop count and maximizing non-blocking bandwidth for collective operations.
A security team requires that the BlueField-3 DPU's embedded Arm subsystem independently own and enforce all NIC configuration and network policy, so that the compromised or untrusted x86 host cannot alter networking behavior. Which BlueField operating mode should be configured on the DPU?
In DPU mode, the embedded Arm subsystem on the BlueField card owns the network functions and control plane. The host sees a virtual NIC managed by the DPU, and privileged network configuration is enforced independently of the host OS, which is exactly what an isolation/zero-trust security requirement calls for.
A cluster uses 400GbE RoCEv2 for GPU-to-GPU NCCL traffic instead of InfiniBand. Engineers observe intermittent packet drops on the switches during large all-reduce operations, causing NCCL collective timeouts. The switches currently have flow control disabled entirely. Which change is MOST likely to resolve the timeouts?
RoCEv2 requires a lossless Ethernet fabric to perform reliably. Enabling PFC on the dedicated RoCE priority prevents buffer overflow drops at the switch, while ECN/WRED provides end-to-end congestion signaling so senders back off before drops occur. Without both mechanisms configured together, RoCEv2 traffic is prone to exactly this kind of drop-induced timeout under load.
An AI training cluster needs to store and retrieve model checkpoints at the highest possible throughput with the lowest latency, and wants to leverage GPUDirect Storage to move data directly between NVMe storage and GPU memory, bypassing the CPU. Which storage connectivity approach should be deployed?
GPUDirect Storage requires an RDMA-capable data path between storage targets and GPU memory. NVMe over Fabrics (NVMe-oF) running over InfiniBand or RoCE provides the low-latency, high-throughput, CPU-bypass connectivity that GDS depends on, making it the standard choice for checkpoint I/O in NVIDIA-validated AI infrastructure designs.
A shared A100 GPU node will host several small, isolated inference workloads that each require guaranteed compute and memory partitions with hardware-level fault isolation. After confirming the GPU supports MIG, which command correctly places the GPU into MIG mode before creating instances?
`nvidia-smi -mig 1` enables MIG mode on the GPU (typically requiring a GPU reset or reboot to take effect). Only after MIG mode is enabled can GPU instances such as 1g.5gb profiles be created with the `-cgi` flag.
You are planning intra-rack InfiniBand NDR 400Gb/s links where some runs exceed 3 meters between a DGX H100 and a leaf switch in an adjacent rack. Which cabling choice is appropriate for this run length at NDR speeds?
At NDR (400Gb/s) speeds, passive copper DAC cables are only reliable for very short reaches (typically around 1.5-2 meters). For runs beyond that distance, such as between adjacent racks, active optical cables or optical transceivers with MPO fiber are required to maintain signal integrity.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$199.99 $34.99 with code FREETEST33 — valid through September 9.
Get my $34.99 deal →