cert
cert-1
cert-2

Pass Cisco 300-640 Exam in First Attempt Guaranteed!

Get 100% Latest Exam Questions, Accurate & Verified Answers to Pass the Actual Exam!
30 Days Free Updates, Instant Download!

cert-5
cert-6
300-640 Exam - Verified By Experts
300-640 Premium File

300-640 Premium File

$89.99
$98.99
  • Premium File 60 Questions & Answers. Last Update: Nov 10, 2026

Whats Included:

  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates
 
$98.99
$89.99
accept 10 downloads in the last 7 days
block-screenshots
Cisco 300-640 Practice Test 9

Question 80

The acceptance board must decide whether an AI-infrastructure response to edge inference resilience is fit for controlled operation. The fictional operating facts are: A remote wind turbine uses a vibration anomaly model but its upstream connection is unavailable for hours during storms. Signoff cannot rely on a good-looking utilization graph, a change ticket alone or an unmeasured promise. The response must include an appropriate governing engineering principle, a scoped intervention and a separate measurable acceptance criterion. Choose EXACTLY THREE statements that meet those roles.

  1. Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits.
  2. Deploy a compact validated inference artifact with signed updates, bounded local buffering and a defined offline fallback.
  3. Workload placement must satisfy binding residency, data-handling and latency constraints before cloud-versus-on-premises cost comparisons.
  4. A deployment recreates the required accelerator placement after node drain without starving protected services.
  5. The edge unit handles a simulated WAN outage and recovers without losing critical safety events.
  6. Correct document permissions and retrieval filters, refresh the stale index and evaluate grounded responses on representative queries.

Explanation

Correct Answer

A:

A remote wind turbine uses a vibration anomaly model but its upstream connection is unavailable for hours during storms. Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits. Deploy a compact validated inference artifact with signed updates, bounded local buffering and a defined offline fallback.

B:

A remote wind turbine uses a vibration anomaly model but its upstream connection is unavailable for hours during storms. Deploy a compact validated inference artifact with signed updates, bounded local buffering and a defined offline fallback. Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits.

E:

A remote wind turbine uses a vibration anomaly model but its upstream connection is unavailable for hours during storms. The edge unit handles a simulated WAN outage and recovers without losing critical safety events. Deploy a compact validated inference artifact with signed updates, bounded local buffering and a defined offline fallback.

Incorrect Answers

C:

Workload placement must satisfy binding residency, data-handling and latency constraints before cloud-versus-on-premises cost comparisons. It is not the rule that supports the proposed change and acceptance test: A remote wind turbine uses a vibration anomaly model but its upstream connection is unavailable for hours during storms. Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits. Deploy a compact validated inference artifact with signed updates, bounded local buffering and a defined offline fallback.

D:

A deployment recreates the required accelerator placement after node drain without starving protected services. It would not independently demonstrate that the control works: A remote wind turbine uses a vibration anomaly model but its upstream connection is unavailable for hours during storms. The edge unit handles a simulated WAN outage and recovers without losing critical safety events. Deploy a compact validated inference artifact with signed updates, bounded local buffering and a defined offline fallback.

F:

Correct document permissions and retrieval filters, refresh the stale index and evaluate grounded responses on representative queries. It does not implement the governing rule for this approval: A remote wind turbine uses a vibration anomaly model but its upstream connection is unavailable for hours during storms. Deploy a compact validated inference artifact with signed updates, bounded local buffering and a defined offline fallback. The edge unit handles a simulated WAN outage and recovers without losing critical safety events.

Question 81

A data-center AI architecture committee considers gpu interconnect topology awareness in its AI Fundamentals and Applications review. At the affected site: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. The committee is deliberately not selecting a vendor product or applying a live change yet. It needs the foundational reasoning that makes the next design decision defensible. Which ONE engineering principle directly governs this situation? Choose ONE answer.

  1. A hybrid workload needs explicit identity, encrypted transport, consistent model and data versions, and ownership of cross-site failure modes.
  2. Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits.
  3. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.
  4. Workload placement must satisfy binding residency, data-handling and latency constraints before cloud-versus-on-premises cost comparisons.

Explanation

Correct Answer

C:

An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.

Incorrect Answers

A:

A hybrid workload needs explicit identity, encrypted transport, consistent model and data versions, and ownership of cross-site failure modes. It does not govern the condition described: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.

B:

Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits. It does not govern the condition described: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.

D:

Workload placement must satisfy binding residency, data-handling and latency constraints before cloud-versus-on-premises cost comparisons. It does not govern the condition described: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.

Question 82

An infrastructure diagnostician is investigating gpu interconnect topology awareness for the AI Fundamentals and Applications team. The reported incident reads: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Competing teams attribute the symptoms to different layers of the AI stack. Before accepting a hypothesis or touching production settings, the investigator must choose a discriminating set of measurements. Which ONE evidence package is MOST relevant? Choose ONE answer.

  1. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.
  2. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
  3. Record on-device model footprint, local input rates, offline duration, update signatures and safe fallback triggers.
  4. Inspect the job specification, checkpoint lineage, optimizer activity, training-data permissions and the separately measured inference path.

Explanation

Correct Answer

A:

An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.

Incorrect Answers

B:

Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load. That check would not establish the decisive facts for this condition: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.

C:

Record on-device model footprint, local input rates, offline duration, update signatures and safe fallback triggers. That check would not establish the decisive facts for this condition: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.

D:

Inspect the job specification, checkpoint lineage, optimizer activity, training-data permissions and the separately measured inference path. That check would not establish the decisive facts for this condition: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.

Question 83

During a bounded AI-platform change-control review concerning gpu interconnect topology awareness, the site owner reports: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. The proposed work must have a narrow implementation scope, first be validated against the failure evidence and avoid unrelated system-wide tuning. Which ONE intervention should the engineers pilot before any wider rollout? Choose ONE answer.

  1. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.
  2. Deploy a compact validated inference artifact with signed updates, bounded local buffering and a defined offline fallback.
  3. Define hardware-aware scheduling constraints, namespaces and resource quotas with documented preemption and restart behavior.
  4. Establish least-privilege hybrid links, versioned artifact replication and fail-closed handling for stale data or disconnected sites.

Explanation

Correct Answer

A:

An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.

Incorrect Answers

B:

Deploy a compact validated inference artifact with signed updates, bounded local buffering and a defined offline fallback. That change would leave the reported problem unresolved: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

C:

Define hardware-aware scheduling constraints, namespaces and resource quotas with documented preemption and restart behavior. That change would leave the reported problem unresolved: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

D:

Establish least-privilege hybrid links, versioned artifact replication and fail-closed handling for stale data or disconnected sites. That change would leave the reported problem unresolved: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

Question 84

Two independent technical reviewers examine a AI Fundamentals and Applications problem involving gpu interconnect topology awareness. Their shared incident log states: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. One reviewer requires measurements capable of separating the observed problem from neighboring failure modes; the other requires a bounded action that follows those measurements. Choose EXACTLY TWO statements that together supply those two parts.

  1. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.
  2. Inspect site-to-cloud connectivity paths, service identities, synchronization lag, data classification and failure recovery objectives.
  3. Review false-positive and false-negative cost, data quality, stakeholder workflow, baseline metrics and feasible human escalation.
  4. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.
  5. Place restricted data and sensitive inference inside approved boundaries; use external services only with approved redaction and transfer controls.

Explanation

Correct Answer

A:

An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

D:

An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

Incorrect Answers

B:

Inspect site-to-cloud connectivity paths, service identities, synchronization lag, data classification and failure recovery objectives. It would not provide the evidence needed for the paired diagnosis and follow-up: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

C:

Review false-positive and false-negative cost, data quality, stakeholder workflow, baseline metrics and feasible human escalation. It would not provide the evidence needed for the paired diagnosis and follow-up: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

E:

Place restricted data and sensitive inference inside approved boundaries; use external services only with approved redaction and transfer controls. It is not the scoped response supported by the accompanying evidence: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

Question 85

The acceptance board must decide whether an AI-infrastructure response to gpu interconnect topology awareness is fit for controlled operation. The fictional operating facts are: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Signoff cannot rely on a good-looking utilization graph, a change ticket alone or an unmeasured promise. The response must include an appropriate governing engineering principle, a scoped intervention and a separate measurable acceptance criterion. Choose EXACTLY THREE statements that meet those roles.

  1. Define hardware-aware scheduling constraints, namespaces and resource quotas with documented preemption and restart behavior.
  2. AI orchestration should request actual GPUs, memory and storage resources and respect scheduling isolation instead of treating all worker nodes as equivalent.
  3. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.
  4. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.
  5. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.
  6. Data loaders sustain the target sample rate while checkpoints restore correctly under a simulated host loss.

Explanation

Correct Answer

C:

An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.

D:

An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Collective benchmark efficiency increases without violating platform support or GPU memory constraints. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

E:

An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

Incorrect Answers

A:

Define hardware-aware scheduling constraints, namespaces and resource quotas with documented preemption and restart behavior. It does not implement the governing rule for this approval: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

B:

AI orchestration should request actual GPUs, memory and storage resources and respect scheduling isolation instead of treating all worker nodes as equivalent. It is not the rule that supports the proposed change and acceptance test: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

F:

Data loaders sustain the target sample rate while checkpoints restore correctly under a simulated host loss. It would not independently demonstrate that the control works: An image-training job doubles GPU cards in a host but collective synchronization time overwhelms added compute capacity. Collective benchmark efficiency increases without violating platform support or GPU memory constraints. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

Question 86

A data-center AI architecture committee considers gpu interconnect topology awareness in its AI Fundamentals and Applications review. At the affected site: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. The committee is deliberately not selecting a vendor product or applying a live change yet. It needs the foundational reasoning that makes the next design decision defensible. Which ONE engineering principle directly governs this situation? Choose ONE answer.

  1. Reliable AI delivery separates data validation, training, evaluation, artifact registry, deployment, telemetry and rollback with reproducible version identifiers.
  2. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.
  3. Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits.
  4. Select block, file or object access by application consistency, sharing and throughput needs; storage capacity alone does not determine suitability.

Explanation

Correct Answer

B:

A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.

Incorrect Answers

A:

Reliable AI delivery separates data validation, training, evaluation, artifact registry, deployment, telemetry and rollback with reproducible version identifiers. It does not govern the condition described: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.

C:

Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits. It does not govern the condition described: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.

D:

Select block, file or object access by application consistency, sharing and throughput needs; storage capacity alone does not determine suitability. It does not govern the condition described: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.

Question 87

An infrastructure diagnostician is investigating gpu interconnect topology awareness for the AI Fundamentals and Applications team. The reported incident reads: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Competing teams attribute the symptoms to different layers of the AI stack. Before accepting a hypothesis or touching production settings, the investigator must choose a discriminating set of measurements. Which ONE evidence package is MOST relevant? Choose ONE answer.

  1. Inspect site-to-cloud connectivity paths, service identities, synchronization lag, data classification and failure recovery objectives.
  2. Capture read/write pattern, shared file semantics, metadata rate, namespace operations, IOPS and bandwidth measurements.
  3. Inspect the job specification, checkpoint lineage, optimizer activity, training-data permissions and the separately measured inference path.
  4. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.

Explanation

Correct Answer

D:

A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.

Incorrect Answers

A:

Inspect site-to-cloud connectivity paths, service identities, synchronization lag, data classification and failure recovery objectives. That check would not establish the decisive facts for this condition: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.

B:

Capture read/write pattern, shared file semantics, metadata rate, namespace operations, IOPS and bandwidth measurements. That check would not establish the decisive facts for this condition: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.

C:

Inspect the job specification, checkpoint lineage, optimizer activity, training-data permissions and the separately measured inference path. That check would not establish the decisive facts for this condition: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.

Question 88

During a bounded AI-platform change-control review concerning gpu interconnect topology awareness, the site owner reports: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. The proposed work must have a narrow implementation scope, first be validated against the failure evidence and avoid unrelated system-wide tuning. Which ONE intervention should the engineers pilot before any wider rollout? Choose ONE answer.

  1. Pilot the narrowly scoped inference workflow with a named owner and benchmark it against a simple deterministic baseline.
  2. Match the training and checkpoint workflow to supported file or block protocols and benchmark access under representative concurrency.
  3. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.
  4. Gate model promotion on evaluation and policy checks, retain lineage and provide a tested rollback to the preceding known-good version.

Explanation

Correct Answer

C:

A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.

Incorrect Answers

A:

Pilot the narrowly scoped inference workflow with a named owner and benchmark it against a simple deterministic baseline. That change would leave the reported problem unresolved: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

B:

Match the training and checkpoint workflow to supported file or block protocols and benchmark access under representative concurrency. That change would leave the reported problem unresolved: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

D:

Gate model promotion on evaluation and policy checks, retain lineage and provide a tested rollback to the preceding known-good version. That change would leave the reported problem unresolved: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

Question 89

Two independent technical reviewers examine a AI Fundamentals and Applications problem involving gpu interconnect topology awareness. Their shared incident log states: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. One reviewer requires measurements capable of separating the observed problem from neighboring failure modes; the other requires a bounded action that follows those measurements. Choose EXACTLY TWO statements that together supply those two parts.

  1. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.
  2. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.
  3. Compare the user's query with retrieved document IDs, access scopes, citation coverage and the freshness of indexed versions.
  4. Record on-device model footprint, local input rates, offline duration, update signatures and safe fallback triggers.
  5. Correct document permissions and retrieval filters, refresh the stale index and evaluate grounded responses on representative queries.

Explanation

Correct Answer

A:

A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

B:

A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

Incorrect Answers

C:

Compare the user's query with retrieved document IDs, access scopes, citation coverage and the freshness of indexed versions. It would not provide the evidence needed for the paired diagnosis and follow-up: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

D:

Record on-device model footprint, local input rates, offline duration, update signatures and safe fallback triggers. It would not provide the evidence needed for the paired diagnosis and follow-up: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

E:

Correct document permissions and retrieval filters, refresh the stale index and evaluate grounded responses on representative queries. It is not the scoped response supported by the accompanying evidence: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

Question 90

The acceptance board must decide whether an AI-infrastructure response to gpu interconnect topology awareness is fit for controlled operation. The fictional operating facts are: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Signoff cannot rely on a good-looking utilization graph, a change ticket alone or an unmeasured promise. The response must include an appropriate governing engineering principle, a scoped intervention and a separate measurable acceptance criterion. Choose EXACTLY THREE statements that meet those roles.

  1. A deployment recreates the required accelerator placement after node drain without starving protected services.
  2. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.
  3. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.
  4. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.
  5. A hybrid workload needs explicit identity, encrypted transport, consistent model and data versions, and ownership of cross-site failure modes.
  6. Match the training and checkpoint workflow to supported file or block protocols and benchmark access under representative concurrency.

Explanation

Correct Answer

B:

A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

C:

A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone.

D:

A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Collective benchmark efficiency increases without violating platform support or GPU memory constraints. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

Incorrect Answers

A:

A deployment recreates the required accelerator placement after node drain without starving protected services. It would not independently demonstrate that the control works: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Collective benchmark efficiency increases without violating platform support or GPU memory constraints. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

E:

A hybrid workload needs explicit identity, encrypted transport, consistent model and data versions, and ownership of cross-site failure modes. It is not the rule that supports the proposed change and acceptance test: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Multi-GPU training design must match collective-communication demands to supported local GPU links, topology and NIC placement rather than count GPU cards alone. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.

F:

Match the training and checkpoint workflow to supported file or block protocols and benchmark access under representative concurrency. It does not implement the governing rule for this approval: A recommendation-model experiment spans accelerators across two CPU sockets and reports irregular all-reduce stalls despite spare FLOPS. Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes. Collective benchmark efficiency increases without violating platform support or GPU memory constraints.

Implementing Cisco Data Center AI Infrastructure (DCAI) (300-640) Practice Tests