8 of the 15 questions in this set, with the correct answer marked and every option explained.
1. What does RoCE provide on a Spectrum-X Ethernet fabric?
✓RDMA semantics over Ethernet, allowing direct memory-to-memory transfers between nodes without involving the host CPU on either side.
RDMA over Converged Ethernet brings InfiniBand-style remote direct memory access to Ethernet fabrics. Removing CPU involvement and buffer copies from the data path is what delivers the low latency and high throughput that distributed training collectives need.
✗Encryption of all traffic between GPUs using hardware offload.
Encryption is a separate capability. RoCE is about the transport semantics — direct memory access — not confidentiality.
✗Automatic compression of gradient tensors before transmission.
Gradient compression is a framework-level technique implemented in software such as NCCL or the training library, not a property of RoCE.
✗Conversion of Ethernet frames into InfiniBand packets for gateway interoperability.
That describes a gateway function. RoCE runs RDMA natively over Ethernet; it does not translate between fabric types.
2. Why are ECN and PFC important on a RoCE fabric carrying AI training traffic?
✓RDMA performs poorly with packet loss, so congestion must be signalled (ECN) and flow controlled (PFC) to keep the fabric effectively lossless under load.
RDMA transports assume a near-lossless network; drops trigger expensive recovery that devastates collective performance. Explicit congestion notification marks packets so senders slow down, and priority flow control pauses a specific traffic class at the link level rather than dropping it.
✗They encrypt congestion telemetry so tenants cannot observe each other's traffic.
Neither mechanism performs encryption. Tenant isolation is addressed by BGP-EVPN and multi-tenancy configuration.
✗They increase the MTU automatically when queues fill.
MTU is a configured property. Congestion control adjusts sending rate and applies backpressure; it does not resize frames.
✗They are only required on the management network, not the compute fabric.
The reverse is true. The compute fabric carries the RDMA traffic that depends on lossless behaviour; the management network does not.
3. What is the purpose of NVIDIA What Just Happened (WJH)?
✓It reports the specific reason individual packets were dropped, turning an opaque counter into an actionable event with context.
Traditional switches tell you a drop counter incremented but not why. WJH provides per-event detail — the reason, the flow, the port — which is what makes congestion and packet-loss diagnosis tractable on a large fabric.
✗It replays historical traffic for capacity planning simulations.
Simulation and what-if modelling is what NVIDIA Air provides. WJH reports real events that already occurred.
✗It automatically reconfigures switch routing tables when congestion is detected.
Automatic path adjustment is adaptive routing. WJH is a telemetry and diagnostics service, not a control mechanism.
✗It manages firmware upgrades across the Spectrum switch fleet.
Fleet firmware management is a separate operational task. WJH concerns packet-level visibility.
4. A service provider must isolate multiple tenants' workloads on shared Spectrum-X switching. Which technology does the blueprint name for this?
✓BGP-EVPN
BGP Ethernet VPN provides the control plane for overlay-based multi-tenancy, keeping each tenant's Layer 2 and Layer 3 domains separate across a shared underlay. The blueprint names configuring multi-tenancy BGP-EVPN to isolate tenant workloads explicitly.
✗PFC
Priority flow control manages congestion per traffic class at the link level. It shapes behaviour, not tenant boundaries.
✗NetQ
NetQ provides real-time network monitoring and validation. It observes the fabric rather than providing tenant isolation.
✗ibdiagnet
ibdiagnet is an InfiniBand diagnostic tool and does not apply to Ethernet multi-tenancy at all.
5. What is NVIDIA Air used for?
✓Simulating network environments so configurations and topologies can be validated before deployment on physical hardware.
Air provides a digital twin of the network, letting engineers build and test a topology and its configuration in software. Catching a design or configuration error there is far cheaper than discovering it after cabling a production fabric.
✗Cooling management for switch chassis in high-density racks.
The name invites this misreading, but Air is a network simulation platform. Thermal management is a facilities concern.
✗Wireless networking for data center management access.
Air has nothing to do with wireless. It simulates wired data center fabrics.
✗Over-the-air firmware distribution to switches.
Firmware distribution is handled through management tooling, not through Air.
6. What is the function of partition keys (PKeys) in an InfiniBand fabric?
✓They logically partition the fabric so that only nodes sharing a partition can communicate, providing multi-tenant isolation.
PKeys are InfiniBand's isolation primitive, roughly analogous to VLANs on Ethernet. Ports are assigned to partitions and traffic is only permitted within a shared partition, which is how secure multi-tenancy is achieved on a shared fabric.
✗They are cryptographic keys that encrypt payloads between endpoints.
PKeys are identifiers for partition membership, not encryption material. They provide segmentation, not confidentiality.
✗They determine the physical port speed negotiated on each link.
Link rate negotiation is a physical layer matter driven by the hardware and cabling, unrelated to partitioning.
✗They index routing tables in the subnet manager for shortest-path calculation.
Routing is computed by the subnet manager using LIDs and topology. PKeys govern who may talk to whom, not path selection.
7. Which platform does the blueprint name for monitoring InfiniBand link status, bandwidth utilization and fabric health?
✓NVIDIA Unified Fabric Manager (UFM)
UFM is the management and monitoring platform for InfiniBand fabrics, covering link state, utilization, health diagnostics and topology awareness. The blueprint references it for both monitoring and system health diagnosis.
✗NetQ
NetQ is the monitoring and validation platform on the Ethernet side, for Cumulus Linux and Spectrum switches. The question specifically asks about InfiniBand.
✗DCGM
DCGM monitors GPUs, not network fabrics.
✗Base Command Manager
BCM manages cluster nodes, images and scheduling. It is not the InfiniBand fabric manager.
8. What problem does adaptive routing solve on a high-performance fabric?
✓It dynamically shifts traffic onto alternative paths in response to congestion, rather than pinning each flow to one statically chosen route.
With static routing, several large flows can hash onto the same path and congest it while parallel paths sit idle. Adaptive routing observes congestion and redistributes traffic across available paths, which matters greatly for the bursty, synchronized collectives of distributed training.
✗It reduces the number of physical cables required between switches.
Adaptive routing makes better use of the paths that exist; it does not reduce the topology's cabling requirements.
✗It encrypts traffic differently on each path to prevent traffic analysis.
No encryption is involved. Adaptive routing is a path selection mechanism.
✗It converts unicast traffic into multicast to reduce total bytes on the wire.
That would be a collective optimization such as in-network aggregation, a distinct capability from adaptive routing.
7 more questions in the app
Practise the full 15-question set with a timer, scoring and progress tracking.