Cisco 300-640 Practice Test 3
Question 20
The acceptance board must decide whether an AI-infrastructure response to separating training from serving is fit for controlled operation. The fictional operating facts are: A biomedical research team reports that a production model's behavior changed overnight even though the deployment tag stayed constant and training jobs shared its artifact directory. 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.
- Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits.
- Keep approved dataset curation and checkpoint promotion in a versioned training workflow and deploy a signed inference artifact separately.
- A reproducible checkpoint version is promoted once and serving requests leave the training dataset unchanged.
- Compare the planned cluster with the current validated Cisco reference architecture and test deviations in a controlled pilot.
- Model training modifies learned parameters using a controlled training pipeline; serving a model for inference should not silently retrain it on each user request.
- A staged test proves authorized transfer, version parity and the documented loss-of-connectivity behavior.
Explanation
Correct Answer
B:A biomedical research team reports that a production model's behavior changed overnight even though the deployment tag stayed constant and training jobs shared its artifact directory. Keep approved dataset curation and checkpoint promotion in a versioned training workflow and deploy a signed inference artifact separately. Model training modifies learned parameters using a controlled training pipeline; serving a model for inference should not silently retrain it on each user request.
C:A biomedical research team reports that a production model's behavior changed overnight even though the deployment tag stayed constant and training jobs shared its artifact directory. A reproducible checkpoint version is promoted once and serving requests leave the training dataset unchanged. Keep approved dataset curation and checkpoint promotion in a versioned training workflow and deploy a signed inference artifact separately.
E:A biomedical research team reports that a production model's behavior changed overnight even though the deployment tag stayed constant and training jobs shared its artifact directory. Model training modifies learned parameters using a controlled training pipeline; serving a model for inference should not silently retrain it on each user request. Keep approved dataset curation and checkpoint promotion in a versioned training workflow and deploy a signed inference artifact separately.
Incorrect Answers
A:Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits. It is not the rule that supports the proposed change and acceptance test: A biomedical research team reports that a production model's behavior changed overnight even though the deployment tag stayed constant and training jobs shared its artifact directory. Model training modifies learned parameters using a controlled training pipeline; serving a model for inference should not silently retrain it on each user request. Keep approved dataset curation and checkpoint promotion in a versioned training workflow and deploy a signed inference artifact separately.
D:Compare the planned cluster with the current validated Cisco reference architecture and test deviations in a controlled pilot. It does not implement the governing rule for this approval: A biomedical research team reports that a production model's behavior changed overnight even though the deployment tag stayed constant and training jobs shared its artifact directory. Keep approved dataset curation and checkpoint promotion in a versioned training workflow and deploy a signed inference artifact separately. A reproducible checkpoint version is promoted once and serving requests leave the training dataset unchanged.
F:A staged test proves authorized transfer, version parity and the documented loss-of-connectivity behavior. It would not independently demonstrate that the control works: A biomedical research team reports that a production model's behavior changed overnight even though the deployment tag stayed constant and training jobs shared its artifact directory. A reproducible checkpoint version is promoted once and serving requests leave the training dataset unchanged. Keep approved dataset curation and checkpoint promotion in a versioned training workflow and deploy a signed inference artifact separately.
Question 21
A data-center AI architecture committee considers interactive inference slos in its AI Fundamentals and Applications review. At the affected site: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. 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.
- Model training modifies learned parameters using a controlled training pipeline; serving a model for inference should not silently retrain it on each user request.
- A retrieval-augmented assistant should fetch authorized current passages and ground its responses in those passages rather than rely on model memory alone.
- Select block, file or object access by application consistency, sharing and throughput needs; storage capacity alone does not determine suitability.
- Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
Explanation
Correct Answer
D:A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
Incorrect Answers
A:Model training modifies learned parameters using a controlled training pipeline; serving a model for inference should not silently retrain it on each user request. It does not govern the condition described: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
B:A retrieval-augmented assistant should fetch authorized current passages and ground its responses in those passages rather than rely on model memory alone. It does not govern the condition described: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
C: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 fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
Question 22
An infrastructure diagnostician is investigating interactive inference slos for the AI Fundamentals and Applications team. The reported incident reads: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. 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.
- Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
- Record on-device model footprint, local input rates, offline duration, update signatures and safe fallback triggers.
- Capture read/write pattern, shared file semantics, metadata rate, namespace operations, IOPS and bandwidth measurements.
- Compare the user's query with retrieved document IDs, access scopes, citation coverage and the freshness of indexed versions.
Explanation
Correct Answer
A:A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
Incorrect Answers
B: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: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
C: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 fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
D:Compare the user's query with retrieved document IDs, access scopes, citation coverage and the freshness of indexed versions. That check would not establish the decisive facts for this condition: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
Question 23
During a bounded AI-platform change-control review concerning interactive inference slos, the site owner reports: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. 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.
- Deploy a compact validated inference artifact with signed updates, bounded local buffering and a defined offline fallback.
- Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
- Gate model promotion on evaluation and policy checks, retain lineage and provide a tested rollback to the preceding known-good version.
- Keep approved dataset curation and checkpoint promotion in a versioned training workflow and deploy a signed inference artifact separately.
Explanation
Correct Answer
B:A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
Incorrect Answers
A: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: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
C: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 fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
D:Keep approved dataset curation and checkpoint promotion in a versioned training workflow and deploy a signed inference artifact separately. That change would leave the reported problem unresolved: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
Question 24
Two independent technical reviewers examine a AI Fundamentals and Applications problem involving interactive inference slos. Their shared incident log states: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. 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.
- Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
- Review false-positive and false-negative cost, data quality, stakeholder workflow, baseline metrics and feasible human escalation.
- Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.
- Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
- Correct document permissions and retrieval filters, refresh the stale index and evaluate grounded responses on representative queries.
Explanation
Correct Answer
A:A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
D:A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
Incorrect Answers
B: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: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
C:Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix. It would not provide the evidence needed for the paired diagnosis and follow-up: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
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 fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
Question 25
The acceptance board must decide whether an AI-infrastructure response to interactive inference slos is fit for controlled operation. The fictional operating facts are: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. 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.
- Workload placement must satisfy binding residency, data-handling and latency constraints before cloud-versus-on-premises cost comparisons.
- Establish least-privilege hybrid links, versioned artifact replication and fail-closed handling for stale data or disconnected sites.
- The edge unit handles a simulated WAN outage and recovers without losing critical safety events.
- Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
- The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
- Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
Explanation
Correct Answer
D:A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
E:A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
F:A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
Incorrect Answers
A: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 fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
B:Establish least-privilege hybrid links, versioned artifact replication and fail-closed handling for stale data or disconnected sites. It does not implement the governing rule for this approval: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
C:The edge unit handles a simulated WAN outage and recovers without losing critical safety events. It would not independently demonstrate that the control works: A fraud-screening API meets average latency in a quiet lab but exceeds its p99 decision deadline during morning account traffic. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
Question 26
A data-center AI architecture committee considers interactive inference slos in its AI Fundamentals and Applications review. At the affected site: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. 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.
- A vendor reference design is a starting constraint set; validate the exact Cisco AI POD or fabric bill of materials, software support and workload fit before implementation.
- Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
- Select block, file or object access by application consistency, sharing and throughput needs; storage capacity alone does not determine suitability.
- Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits.
Explanation
Correct Answer
B:A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
Incorrect Answers
A:A vendor reference design is a starting constraint set; validate the exact Cisco AI POD or fabric bill of materials, software support and workload fit before implementation. It does not govern the condition described: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
C: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 multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
D:Edge AI should tolerate intermittent upstream connectivity while honoring local compute, power, update and privacy limits. It does not govern the condition described: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
Question 27
An infrastructure diagnostician is investigating interactive inference slos for the AI Fundamentals and Applications team. The reported incident reads: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. 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.
- Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
- Record on-device model footprint, local input rates, offline duration, update signatures and safe fallback triggers.
- Inspect site-to-cloud connectivity paths, service identities, synchronization lag, data classification and failure recovery objectives.
- Inspect the job specification, checkpoint lineage, optimizer activity, training-data permissions and the separately measured inference path.
Explanation
Correct Answer
A:A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
Incorrect Answers
B: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: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
C: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 multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
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: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
Question 28
During a bounded AI-platform change-control review concerning interactive inference slos, the site owner reports: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. 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.
- Gate model promotion on evaluation and policy checks, retain lineage and provide a tested rollback to the preceding known-good version.
- Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
- Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes.
- Match the training and checkpoint workflow to supported file or block protocols and benchmark access under representative concurrency.
Explanation
Correct Answer
B:A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
Incorrect Answers
A: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 multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
C:Place tightly coupled workers on topology-compatible GPU groups and validate communication scaling before expanding nodes. That change would leave the reported problem unresolved: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
D: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 multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
Question 29
Two independent technical reviewers examine a AI Fundamentals and Applications problem involving interactive inference slos. Their shared incident log states: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. 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.
- Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix.
- Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
- Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load.
- Inspect pod resource requests, device-plugin visibility, node labels, quota, volume readiness and scheduling events.
- Compare the planned cluster with the current validated Cisco reference architecture and test deviations in a controlled pilot.
Explanation
Correct Answer
B:A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
C:A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
Incorrect Answers
A:Measure all-reduce duration, PCIe and NVLink topology, GPU locality, host-to-network contention and the supported platform matrix. It would not provide the evidence needed for the paired diagnosis and follow-up: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
D:Inspect pod resource requests, device-plugin visibility, node labels, quota, volume readiness and scheduling events. It would not provide the evidence needed for the paired diagnosis and follow-up: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Measure per-token or per-request p95 and p99 latency, queue wait, active batch size, concurrency and GPU memory headroom under load. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
E:Compare the planned cluster with the current validated Cisco reference architecture and test deviations in a controlled pilot. It is not the scoped response supported by the accompanying evidence: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
Question 30
The acceptance board must decide whether an AI-infrastructure response to interactive inference slos is fit for controlled operation. The fictional operating facts are: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. 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.
- A staged test proves authorized transfer, version parity and the documented loss-of-connectivity behavior.
- Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
- Establish least-privilege hybrid links, versioned artifact replication and fail-closed handling for stale data or disconnected sites.
- Choose an AI use case only after defining the decision, reliable labels or feedback, error costs, human controls and a non-AI baseline.
- The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
- Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
Explanation
Correct Answer
B:A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
E:A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
F:A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization.
Incorrect Answers
A:A staged test proves authorized transfer, version parity and the documented loss-of-connectivity behavior. It would not independently demonstrate that the control works: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.
C:Establish least-privilege hybrid links, versioned artifact replication and fail-closed handling for stale data or disconnected sites. It does not implement the governing rule for this approval: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective. The approved peak test meets p99 response latency and error-rate budgets without exhausting accelerator memory.
D:Choose an AI use case only after defining the decision, reliable labels or feedback, error costs, human controls and a non-AI baseline. It is not the rule that supports the proposed change and acceptance test: A multilingual chatbot rolls out larger batches and reports higher GPU throughput while its busiest customer conversations regularly time out. Latency-sensitive inference capacity must be sized for peak request arrival, model memory, batching and tail response time, not merely average accelerator utilization. Separate interactive serving capacity from bulk training work and tune batching or autoscaling against an explicit tail-latency objective.