Cisco 300-640 Practice Test 57
Question 560
The acceptance board must decide whether an AI-infrastructure response to tracing rdma path performance is fit for controlled operation. The fictional operating facts are: An AI storage client shows RDMA completions timing out only through one uplink with an unexpected MTU setting. 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.
- Cross-system logs align within tolerance and scheduled job authentication succeeds.
- Server health incidents need correlated power, thermal, adapter and inventory evidence before blaming the AI model.
- Repair the measured inconsistency on the affected path and retest with the same RDMA workload.
- RDMA performance faults require mapping NIC, switch and remote endpoint telemetry along the actual routed path.
- RDMA completions and latency return to baseline while healthy paths remain unchanged.
- Stop rollout on breached acceptance criteria and restore a tested configuration before root-cause rework.
Explanation
Correct Answer
C:An AI storage client shows RDMA completions timing out only through one uplink with an unexpected MTU setting. Repair the measured inconsistency on the affected path and retest with the same RDMA workload. RDMA performance faults require mapping NIC, switch and remote endpoint telemetry along the actual routed path.
D:An AI storage client shows RDMA completions timing out only through one uplink with an unexpected MTU setting. RDMA performance faults require mapping NIC, switch and remote endpoint telemetry along the actual routed path. Repair the measured inconsistency on the affected path and retest with the same RDMA workload.
E:An AI storage client shows RDMA completions timing out only through one uplink with an unexpected MTU setting. RDMA completions and latency return to baseline while healthy paths remain unchanged. Repair the measured inconsistency on the affected path and retest with the same RDMA workload.
Incorrect Answers
A:Cross-system logs align within tolerance and scheduled job authentication succeeds. It would not independently demonstrate that the control works: An AI storage client shows RDMA completions timing out only through one uplink with an unexpected MTU setting. RDMA completions and latency return to baseline while healthy paths remain unchanged. Repair the measured inconsistency on the affected path and retest with the same RDMA workload.
B:Server health incidents need correlated power, thermal, adapter and inventory evidence before blaming the AI model. It is not the rule that supports the proposed change and acceptance test: An AI storage client shows RDMA completions timing out only through one uplink with an unexpected MTU setting. RDMA performance faults require mapping NIC, switch and remote endpoint telemetry along the actual routed path. Repair the measured inconsistency on the affected path and retest with the same RDMA workload.
F:Stop rollout on breached acceptance criteria and restore a tested configuration before root-cause rework. It does not implement the governing rule for this approval: An AI storage client shows RDMA completions timing out only through one uplink with an unexpected MTU setting. Repair the measured inconsistency on the affected path and retest with the same RDMA workload. RDMA completions and latency return to baseline while healthy paths remain unchanged.
Question 561
A data-center AI architecture committee considers storage latency versus gpu stall attribution in its AI Infrastructure Operations and Troubleshooting review. At the affected site: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. 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.
- Operational changes should have independent service acceptance and a bounded rollback trigger rather than trust a successful automation response.
- Network operations should detect degrading AI fabric conditions using correlated interface, queue and topology signals.
- GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
- Actionable monitoring defines thresholds, ownership, severity and escalation so operators can distinguish normal training bursts from faults.
Explanation
Correct Answer
C:A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
Incorrect Answers
A:Operational changes should have independent service acceptance and a bounded rollback trigger rather than trust a successful automation response. It does not govern the condition described: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
B:Network operations should detect degrading AI fabric conditions using correlated interface, queue and topology signals. It does not govern the condition described: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
D:Actionable monitoring defines thresholds, ownership, severity and escalation so operators can distinguish normal training bursts from faults. It does not govern the condition described: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
Question 562
An infrastructure diagnostician is investigating storage latency versus gpu stall attribution for the AI Infrastructure Operations and Troubleshooting team. The reported incident reads: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. 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.
- Collect RDMA retry or completion data, NIC firmware, path MTU, DSCP mappings and interface counters hop by hop.
- Inspect Intersight equipment faults, fan or PSU telemetry, NIC error counters, component history and service-profile association.
- Measure NTP reachability, UTC offset, source hierarchy, log timestamps and drift during node restarts.
- Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
Explanation
Correct Answer
D:A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
Incorrect Answers
A:Collect RDMA retry or completion data, NIC firmware, path MTU, DSCP mappings and interface counters hop by hop. That check would not establish the decisive facts for this condition: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
B:Inspect Intersight equipment faults, fan or PSU telemetry, NIC error counters, component history and service-profile association. That check would not establish the decisive facts for this condition: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
C:Measure NTP reachability, UTC offset, source hierarchy, log timestamps and drift during node restarts. That check would not establish the decisive facts for this condition: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
Question 563
During a bounded AI-platform change-control review concerning storage latency versus gpu stall attribution, the site owner reports: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. 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.
- Stop rollout on breached acceptance criteria and restore a tested configuration before root-cause rework.
- Restore validated cooling or power headroom after facility review and verify under sustained load.
- Restore supported redundant time synchronization and verify event ordering before further root-cause changes.
- Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
Explanation
Correct Answer
D:A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
Incorrect Answers
A:Stop rollout on breached acceptance criteria and restore a tested configuration before root-cause rework. That change would leave the reported problem unresolved: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
B:Restore validated cooling or power headroom after facility review and verify under sustained load. That change would leave the reported problem unresolved: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
C:Restore supported redundant time synchronization and verify event ordering before further root-cause changes. That change would leave the reported problem unresolved: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
Question 564
Two independent technical reviewers examine a AI Infrastructure Operations and Troubleshooting problem involving storage latency versus gpu stall attribution. Their shared incident log states: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. 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.
- Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
- Inspect Intersight equipment faults, fan or PSU telemetry, NIC error counters, component history and service-profile association.
- Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
- Compare Nexus Dashboard fabric health alerts, interface errors, queue telemetry, topology changes and affected job IDs.
- Restore validated cooling or power headroom after facility review and verify under sustained load.
Explanation
Correct Answer
A:A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy. Training throughput improves and measured input wait falls under representative concurrency.
C:A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity. Training throughput improves and measured input wait falls under representative concurrency.
Incorrect Answers
B:Inspect Intersight equipment faults, fan or PSU telemetry, NIC error counters, component history and service-profile association. It would not provide the evidence needed for the paired diagnosis and follow-up: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy. Training throughput improves and measured input wait falls under representative concurrency.
D:Compare Nexus Dashboard fabric health alerts, interface errors, queue telemetry, topology changes and affected job IDs. It would not provide the evidence needed for the paired diagnosis and follow-up: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy. Training throughput improves and measured input wait falls under representative concurrency.
E:Restore validated cooling or power headroom after facility review and verify under sustained load. It is not the scoped response supported by the accompanying evidence: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity. Training throughput improves and measured input wait falls under representative concurrency.
Question 565
The acceptance board must decide whether an AI-infrastructure response to storage latency versus gpu stall attribution is fit for controlled operation. The fictional operating facts are: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. 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.
- GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
- Training throughput improves and measured input wait falls under representative concurrency.
- Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
- Tune symptom-based alerts and route to responsible teams with runbook-linked evidence.
- Network operations should detect degrading AI fabric conditions using correlated interface, queue and topology signals.
- RDMA completions and latency return to baseline while healthy paths remain unchanged.
Explanation
Correct Answer
A:A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
B:A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Training throughput improves and measured input wait falls under representative concurrency. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
C:A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
Incorrect Answers
D:Tune symptom-based alerts and route to responsible teams with runbook-linked evidence. It does not implement the governing rule for this approval: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity. Training throughput improves and measured input wait falls under representative concurrency.
E:Network operations should detect degrading AI fabric conditions using correlated interface, queue and topology signals. It is not the rule that supports the proposed change and acceptance test: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
F:RDMA completions and latency return to baseline while healthy paths remain unchanged. It would not independently demonstrate that the control works: A language-model trainer spends half each step waiting for file metadata while GPUs appear underutilized. Training throughput improves and measured input wait falls under representative concurrency. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
Question 566
A data-center AI architecture committee considers storage latency versus gpu stall attribution in its AI Infrastructure Operations and Troubleshooting review. At the affected site: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. 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.
- Operational telemetry and certificate validation depend on reliable time across fabric and compute components.
- Actionable monitoring defines thresholds, ownership, severity and escalation so operators can distinguish normal training bursts from faults.
- GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
- Operational changes should have independent service acceptance and a bounded rollback trigger rather than trust a successful automation response.
Explanation
Correct Answer
C:A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
Incorrect Answers
A:Operational telemetry and certificate validation depend on reliable time across fabric and compute components. It does not govern the condition described: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
B:Actionable monitoring defines thresholds, ownership, severity and escalation so operators can distinguish normal training bursts from faults. It does not govern the condition described: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
D:Operational changes should have independent service acceptance and a bounded rollback trigger rather than trust a successful automation response. It does not govern the condition described: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
Question 567
An infrastructure diagnostician is investigating storage latency versus gpu stall attribution for the AI Infrastructure Operations and Troubleshooting team. The reported incident reads: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. 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.
- Audit alert frequency, baselines, on-call response, suppression windows and missed failure examples.
- Compare known-good and canary job metrics, error rates, fabric health and rollback completion time.
- Plot switch queue depth, ECN marks, packet drops, RTT and endpoint rate response over synchronized intervals.
- Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
Explanation
Correct Answer
D:A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
Incorrect Answers
A:Audit alert frequency, baselines, on-call response, suppression windows and missed failure examples. That check would not establish the decisive facts for this condition: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
B:Compare known-good and canary job metrics, error rates, fabric health and rollback completion time. That check would not establish the decisive facts for this condition: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
C:Plot switch queue depth, ECN marks, packet drops, RTT and endpoint rate response over synchronized intervals. That check would not establish the decisive facts for this condition: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
Question 568
During a bounded AI-platform change-control review concerning storage latency versus gpu stall attribution, the site owner reports: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. 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.
- Tune symptom-based alerts and route to responsible teams with runbook-linked evidence.
- Isolate the affected host and apply the documented hardware maintenance procedure without mass model redeployment.
- Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
- Set actionable per-symptom thresholds and investigate the shared fabric path before broad policy changes.
Explanation
Correct Answer
C:A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
Incorrect Answers
A:Tune symptom-based alerts and route to responsible teams with runbook-linked evidence. That change would leave the reported problem unresolved: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
B:Isolate the affected host and apply the documented hardware maintenance procedure without mass model redeployment. That change would leave the reported problem unresolved: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
D:Set actionable per-symptom thresholds and investigate the shared fabric path before broad policy changes. That change would leave the reported problem unresolved: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
Question 569
Two independent technical reviewers examine a AI Infrastructure Operations and Troubleshooting problem involving storage latency versus gpu stall attribution. Their shared incident log states: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. 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.
- Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy.
- Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
- Identify and relieve the root congestion source, then confine PFC and verify queue policy safely.
- Trace per-priority pause duration, ingress queue depth, flow source and blocked unrelated classes across hops.
- Correlate GPU clock rate, power limit flags, inlet temperature, UCS faults and workload intensity.
Explanation
Correct Answer
A:A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy. Training throughput improves and measured input wait falls under representative concurrency.
B:A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity. Training throughput improves and measured input wait falls under representative concurrency.
Incorrect Answers
C:Identify and relieve the root congestion source, then confine PFC and verify queue policy safely. It is not the scoped response supported by the accompanying evidence: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity. Training throughput improves and measured input wait falls under representative concurrency.
D:Trace per-priority pause duration, ingress queue depth, flow source and blocked unrelated classes across hops. It would not provide the evidence needed for the paired diagnosis and follow-up: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy. Training throughput improves and measured input wait falls under representative concurrency.
E:Correlate GPU clock rate, power limit flags, inlet temperature, UCS faults and workload intensity. It would not provide the evidence needed for the paired diagnosis and follow-up: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Compare data-loader wait, storage read p99, metadata latency, network path and GPU kernel occupancy. Training throughput improves and measured input wait falls under representative concurrency.
Question 570
The acceptance board must decide whether an AI-infrastructure response to storage latency versus gpu stall attribution is fit for controlled operation. The fictional operating facts are: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. 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.
- GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
- Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
- Training throughput improves and measured input wait falls under representative concurrency.
- Actionable monitoring defines thresholds, ownership, severity and escalation so operators can distinguish normal training bursts from faults.
- The incident report reconstructs a reproducible causal sequence and a verified corrective action.
- Repair the measured inconsistency on the affected path and retest with the same RDMA workload.
Explanation
Correct Answer
A:A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
B:A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence.
C:A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Training throughput improves and measured input wait falls under representative concurrency. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
Incorrect Answers
D:Actionable monitoring defines thresholds, ownership, severity and escalation so operators can distinguish normal training bursts from faults. It is not the rule that supports the proposed change and acceptance test: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. GPU idle periods may originate in storage or input pipelines, so remediation needs cross-layer timing evidence. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
E:The incident report reconstructs a reproducible causal sequence and a verified corrective action. It would not independently demonstrate that the control works: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Training throughput improves and measured input wait falls under representative concurrency. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity.
F:Repair the measured inconsistency on the affected path and retest with the same RDMA workload. It does not implement the governing rule for this approval: A vision farm reports GPU stalls after new dataset shards move to a lower-throughput storage tier. Correct the evidenced storage or pipeline bottleneck before increasing accelerator capacity. Training throughput improves and measured input wait falls under representative concurrency.