Physical Layer Management
Free NVIDIA-Certified Professional: AI Infrastructure practice — 6 questions on Physical Layer Management, with explanations. No sign-up. Full 12-question mixed test →
A leaf switch in an InfiniBand fat-tree fabric has 40 total ports: 32 connected to compute node NICs and 8 uplinks to spine switches. What is the oversubscription ratio of this leaf switch, and what does it indicate?
32 downlink ports divided by 8 uplink ports yields a 4:1 oversubscription ratio, meaning the fabric is oversubscribed and NCCL-style collective operations (all-reduce, all-gather) that saturate node-to-node bandwidth may experience congestion at the spine layer.
In a rail-optimized fat-tree design for a DGX H100 SuperPOD, each of the 8 GPU NICs (one per rail) on a node connects to a dedicated leaf switch reserved solely for that rail. What is the primary benefit of this design for distributed training workloads?
Rail-optimized topology places same-rail GPUs from different nodes onto the same set of leaf switches, so intra-rail collective communication (e.g., ring/tree NCCL patterns matched to GPU index) stays local to the leaf, reducing hop count and spine congestion.
A GPU cluster's spine switches are located in a separate row of racks 50 meters from the compute racks. The fabric uses NDR (400Gb/s) InfiniBand. Which cabling solution should be used for these inter-row links?
At NDR (400Gb/s), passive DAC is limited to roughly 2 meters and active copper only modestly extends that reach. A 50-meter run requires active optical cables (AOC), which integrate transceivers with the fiber assembly and reliably support distances up to about 100 meters.
During physical-layer troubleshooting of a ConnectX-7 port suspected of marginal signal quality, which command-line utility provides RX/TX optical power levels, transceiver temperature, and per-lane symbol error counts for the specific cable?
mlxlink is the NVIDIA/Mellanox utility designed to query detailed per-port physical-layer diagnostics, including optical power levels, transceiver temperature, FEC/symbol error counters, and eye-margin data for a specific cable and port.
A data center rack is fed by a 3-phase 208V PDU circuit rated at 60A per phase. Each DGX H100 system has a maximum power draw of 10.2kW. Applying the NEC 80% continuous-load derating rule, how many DGX H100 systems can this single circuit safely support?
Available continuous capacity = √3 × 208V × 60A × 0.8 ≈ 17.3kW. A single DGX H100 at 10.2kW fits within that budget, but a second unit would require 20.4kW total, exceeding the derated 17.3kW limit — so only 1 system can be safely powered from this circuit.
Which physical-layer monitoring approach correctly identifies a degrading fiber cable in an InfiniBand fabric before it causes a fatal link failure?
ibdiagnet's cable-info reporting captures per-link symbol error trends and transceiver telemetry such as temperature and optical power, enabling proactive detection of degrading fiber or optics before a hard link failure occurs.
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 →