NVIDIA Spectrum Networking
Free NVIDIA-Certified Professional: AI Networking practice — 6 questions on NVIDIA Spectrum Networking, with explanations. No sign-up. Full 12-question mixed test →
A GPU cluster running a 3-tier Spectrum-X fabric begins experiencing a PFC deadlock during a large all-reduce collective, where pause frames propagate across multiple switches and stall traffic in a cyclic pattern. Which approach BEST resolves this without disabling lossless behavior needed for RoCEv2?
Spectrum-X design relies on ECN/DCQCN to proactively signal congestion at the source before PFC pause is triggered, while PFC storm prevention detects and breaks pathological pause loops (deadlocks) that pure PFC-only designs are prone to in multi-tier fabrics.
A network architect is designing a Spectrum-X fabric and needs the GPU-attached NICs to perform RoCEv2 congestion telemetry offload and packet reordering so that adaptive routing can be used without exposing out-of-order packets to the RDMA application. Which NIC/DPU class is required at the host to support this Spectrum-X capability?
BlueField-3 SuperNICs provide the hardware RoCE transport offload, out-of-order packet reordering, and congestion telemetry needed to work with Spectrum-4 switches' adaptive routing while presenting in-order delivery to the GPU application.
On a Spectrum switch configured for RoCEv2 with DCQCN, which set of parameters directly controls the probability that a packet's ECN bit is marked as the egress queue depth grows, before any PFC pause is triggered?
WRED (Weighted Random Early Detection) thresholds Kmin, Kmax, and Pmax on the switch determine at what queue occupancy ECN marking begins, how it scales, and the maximum marking probability — this is the core mechanism feeding DCQCN's congestion feedback loop.
Spectrum-X fabrics use per-packet adaptive routing to maximize path utilization across spine-leaf topologies for RoCEv2 traffic. Why does this NOT cause the out-of-order delivery problems normally associated with per-packet load balancing on RDMA fabrics?
The BlueField-3 SuperNIC includes a hardware reordering engine that reassembles packets received out of order due to adaptive per-packet spraying, presenting an in-order RDMA stream to the application while still gaining the load-balancing benefits of adaptive routing.
During a large language model training run, engineers observe sporadic tail-latency spikes on specific GPU-to-GPU flows but see no obvious link errors or interface drops in standard counters. Which Spectrum switch capability should they use to get real-time, per-event visibility into the exact cause (e.g., buffer congestion, ECN marking, or micro-burst drops) without impacting production traffic?
What Just Happened (WJH) is a Spectrum switch hardware telemetry feature that streams detailed, real-time root-cause events (buffer drops, ECN marks, latency spikes, ACL drops, etc.) without sampling gaps or added traffic load, making it ideal for diagnosing intermittent micro-burst issues.
An architect is calculating PFC headroom buffer requirements for links between Spectrum switches and BlueField SuperNICs where the fiber cable run is being extended from 3 meters to 100 meters. How must the headroom buffer allocation change to prevent packet loss during the pause frame propagation delay?
Headroom buffer must account for all bytes already transmitted before a PFC pause frame is received and takes effect, which is a function of round-trip propagation delay; longer cables mean more in-flight bytes at line rate, requiring larger headroom buffers to avoid drops.
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.
undefined $34.99 with code FREETEST34 — valid through Oct 11.
Get my $34.99 deal →