Free 300-640 practice test: 30 questions with answers

AI back-end fabrics need RDMA with lossless behaviour, provided by RoCE with congestion control or by InfiniBand. Best-effort Ethernet drops packets under load, and dial-up or serial links lack the bandwidth.

Leaf-spine designs offer consistent paths and scale horizontally, which suits east-west AI traffic. Single links and removing switches are not design goals, and no topology eliminates latency.

ECN and DCQCN-style congestion control signal senders to slow down before buffers overflow, which keeps the fabric lossless. Root guard, DHCP snooping, and syslog filtering do not manage congestion.

PFC pauses a specific priority class so RDMA traffic is not dropped. Portfast, sticky MAC learning, and DHCP relay do not provide lossless traffic classes.

Separate fabrics or isolated classes and VRFs keep the sensitive back-end traffic from interfering with general traffic. Mixing everything, disabling the back-end, and single firewall rules do not provide that isolation.

Nexus switches with NX-OS are the common data centre platform for leaf-spine fabrics. Aironet access points, home switches, and branch routers are not data centre fabric platforms.

Comparing error and congestion counters before and after the firmware change identifies network regressions. Monitor restarts, renaming jobs, and reinstalling drivers without checking the network do not isolate the cause.

Switch counters and queue statistics show congestion in the fabric, and NIC and GPU telemetry show communication timing. Wallpapers, toner levels, and browser caches are unrelated.

Interface, VLAN, and MTU settings are common causes of storage connectivity problems after a change. Wallpapers, names, and removing GPUs do not address the connectivity.

Optics, cabling, and error counters on both ends commonly reveal the cause of flapping links. Rebooting GPUs, changing monitors, and reinstalling operating systems do not address link health.

Congestion notification counters and queue depth show whether packets are being marked or queued. File size, user counts, and office temperature do not indicate network congestion.

Sustained low utilization relative to peers suggests the node is idle or starved, which points to scheduling, data, or network issues. The other options are not supported by the pattern.

NVLink and NVSwitch provide high-bandwidth GPU-to-GPU communication inside the server. Wi-Fi, serial sharing, and USB 1.1 cannot match that bandwidth.

Burn-in and benchmarks with the expected workload validate real performance and stability before production. Joining untested, relying on specifications, and idle-only tests do not confirm behaviour under load.

UCS servers support GPUs and are managed through Intersight. Wireless controllers, home switches, and cameras are not AI compute platforms.

Intersight provides cloud-based management for UCS and other Cisco infrastructure. Packet Tracer is a learning simulator, Webex is collaboration software, and Umbrella is DNS security.

A power policy aligned to the GPU requirements prevents power-related throttling. Screen brightness, cable length, and memory size do not control GPU power.

Dense GPU systems produce heavy thermal loads, and inadequate cooling leads to throttling or hardware failures. The other statements are incorrect.

Templates or infrastructure as code with validation deliver consistent, reviewable configuration at scale. Manual memory-based configuration, unreviewed copies, and reactive changes produce drift.

GPU requests and limits in the pod spec let the scheduler allocate GPUs. Printer labels, DNS records, and banners do not allocate GPU resources.

Slurm schedules and manages jobs on compute resources, including GPUs. Packet capture, printing, and web serving are different functions.

Intent-based configuration with reconciliation and drift detection keeps the running state aligned with the declared intent. Sticky notes, accepted drift, and live-only edits break traceability.

An orchestration workflow calls network and compute APIs in the right order so that resources match. Manual, uncoordinated, and naming-based approaches create mismatches.

The same versioned templates with parameters keep environments consistent and promotable. Separate builds, undocumented settings, and uncontrolled data copies cause drift and risk.

GPU starvation from slow data reads is best addressed by improving sustained parallel read throughput. Colour, labels, and folder counts do not affect data delivery.

Parallel or distributed file systems serve many GPU nodes concurrently with high throughput. Local-only disks, tape, and printing do not meet parallel training read patterns.

High-throughput storage and a sensible checkpoint interval reduce write overhead while protecting progress. Slow archives every second, no checkpoints, and memory-only storage either slow the job or risk losing it.

Replication or erasure coding protects data against device failure. A single laptop copy, printed copies, and temporary folders do not provide durable protection.

Hot data on fast nearby storage reduces latency, while colder data can be tiered. Remote placement, tape for hot data, and no local data increase latency.

Dataset size and growth, plus read throughput needed to keep GPUs busy, determine storage sizing. Colour, stapler brands, and coffee breaks are unrelated.