Cisco 300-640 Practice Test 22
Question 210
The acceptance board must decide whether an AI-infrastructure response to compute and fabric lifecycle compatibility is fit for controlled operation. The fictional operating facts are: An AI platform's container image assumes newer device interfaces than a subset of older compute nodes provide. 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.
- AI servers, firmware, NIC drivers, GPU stack and fabric software form a versioned support matrix.
- The staged cluster meets interoperability tests before production promotion.
- The same-load measurement demonstrates reduced facility overhead without harming service performance.
- Training, management, tenant and sensitive data paths need enforceable policy boundaries independent of aggregate throughput.
- Compare like-for-like PUE baselines and optimize cooling or capacity with a documented energy measurement plan.
- Validate and stage a supported end-to-end software bill of materials with rollback images.
Explanation
Correct Answer
A:An AI platform's container image assumes newer device interfaces than a subset of older compute nodes provide. AI servers, firmware, NIC drivers, GPU stack and fabric software form a versioned support matrix. Validate and stage a supported end-to-end software bill of materials with rollback images.
B:An AI platform's container image assumes newer device interfaces than a subset of older compute nodes provide. The staged cluster meets interoperability tests before production promotion. Validate and stage a supported end-to-end software bill of materials with rollback images.
F:An AI platform's container image assumes newer device interfaces than a subset of older compute nodes provide. Validate and stage a supported end-to-end software bill of materials with rollback images. AI servers, firmware, NIC drivers, GPU stack and fabric software form a versioned support matrix.
Incorrect Answers
C:The same-load measurement demonstrates reduced facility overhead without harming service performance. It would not independently demonstrate that the control works: An AI platform's container image assumes newer device interfaces than a subset of older compute nodes provide. The staged cluster meets interoperability tests before production promotion. Validate and stage a supported end-to-end software bill of materials with rollback images.
D:Training, management, tenant and sensitive data paths need enforceable policy boundaries independent of aggregate throughput. It is not the rule that supports the proposed change and acceptance test: An AI platform's container image assumes newer device interfaces than a subset of older compute nodes provide. AI servers, firmware, NIC drivers, GPU stack and fabric software form a versioned support matrix. Validate and stage a supported end-to-end software bill of materials with rollback images.
E:Compare like-for-like PUE baselines and optimize cooling or capacity with a documented energy measurement plan. It does not implement the governing rule for this approval: An AI platform's container image assumes newer device interfaces than a subset of older compute nodes provide. Validate and stage a supported end-to-end software bill of materials with rollback images. The staged cluster meets interoperability tests before production promotion.
Question 211
A data-center AI architecture committee considers resilient interactive service capacity in its AI Infrastructure Components and Architecture review. At the affected site: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. 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.
- Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
- A low-average-latency network can still fail serving SLOs when transient queuing creates high p99 latency.
- Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
- Training, management, tenant and sensitive data paths need enforceable policy boundaries independent of aggregate throughput.
Explanation
Correct Answer
C:A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
Incorrect Answers
A:Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies. It does not govern the condition described: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
B:A low-average-latency network can still fail serving SLOs when transient queuing creates high p99 latency. It does not govern the condition described: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
D:Training, management, tenant and sensitive data paths need enforceable policy boundaries independent of aggregate throughput. It does not govern the condition described: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
Question 212
An infrastructure diagnostician is investigating resilient interactive service capacity for the AI Infrastructure Components and Architecture team. The reported incident reads: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. 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.
- Inventory immutable datasets, intermediate artifacts, checkpoint rotations, compression and quota high-water marks.
- Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
- Compare dataset manifests, content hashes, model versions, replication lag and divergent write histories.
- Collect out-of-memory traces, per-rank allocator usage, peak activations, tensor sizes and chosen precision.
Explanation
Correct Answer
B:A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
Incorrect Answers
A:Inventory immutable datasets, intermediate artifacts, checkpoint rotations, compression and quota high-water marks. That check would not establish the decisive facts for this condition: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
C:Compare dataset manifests, content hashes, model versions, replication lag and divergent write histories. That check would not establish the decisive facts for this condition: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
D:Collect out-of-memory traces, per-rank allocator usage, peak activations, tensor sizes and chosen precision. That check would not establish the decisive facts for this condition: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
Question 213
During a bounded AI-platform change-control review concerning resilient interactive service capacity, the site owner reports: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. 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.
- Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
- Adopt only vendor-supported accelerator-sharing modes and reserve dedicated resources for jobs that need full-device behavior.
- Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported.
- Revise supported precision, checkpointing or model sharding and select GPU memory capacity based on peak use.
Explanation
Correct Answer
A:A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
Incorrect Answers
B:Adopt only vendor-supported accelerator-sharing modes and reserve dedicated resources for jobs that need full-device behavior. That change would leave the reported problem unresolved: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
C:Use workload-aware anti-affinity and recovery checkpoints across independent domains where supported. That change would leave the reported problem unresolved: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
D:Revise supported precision, checkpointing or model sharding and select GPU memory capacity based on peak use. That change would leave the reported problem unresolved: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
Question 214
Two independent technical reviewers examine a AI Infrastructure Components and Architecture problem involving resilient interactive service capacity. Their shared incident log states: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. 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.
- Adopt only vendor-supported accelerator-sharing modes and reserve dedicated resources for jobs that need full-device behavior.
- Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
- Check checkpoint frequency, write completion, replicas, integrity hashes and actual full-state restore time.
- Record adapter firmware, host kernel, accelerator driver, CUDA/runtime level, Cisco UCS profiles and network release compatibility.
- Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
Explanation
Correct Answer
B:A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
E:A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
Incorrect Answers
A:Adopt only vendor-supported accelerator-sharing modes and reserve dedicated resources for jobs that need full-device behavior. It is not the scoped response supported by the accompanying evidence: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
C:Check checkpoint frequency, write completion, replicas, integrity hashes and actual full-state restore time. It would not provide the evidence needed for the paired diagnosis and follow-up: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
D:Record adapter firmware, host kernel, accelerator driver, CUDA/runtime level, Cisco UCS profiles and network release compatibility. It would not provide the evidence needed for the paired diagnosis and follow-up: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
Question 215
The acceptance board must decide whether an AI-infrastructure response to resilient interactive service capacity is fit for controlled operation. The fictional operating facts are: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. 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 controlled replica-loss drill retains the peak request target and bounded p99 latency.
- Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
- Adopt only vendor-supported accelerator-sharing modes and reserve dedicated resources for jobs that need full-device behavior.
- A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives.
- Distributed training requires adequate effective east-west throughput between GPU workers, with oversubscription based on real collective traffic.
- Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
Explanation
Correct Answer
A:A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. A controlled replica-loss drill retains the peak request target and bounded p99 latency. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
B:A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
F:A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
Incorrect Answers
C:Adopt only vendor-supported accelerator-sharing modes and reserve dedicated resources for jobs that need full-device behavior. It does not implement the governing rule for this approval: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
D:A planned node or rack loss restarts the job from an approved checkpoint within recovery objectives. It would not independently demonstrate that the control works: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. A controlled replica-loss drill retains the peak request target and bounded p99 latency. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
E:Distributed training requires adequate effective east-west throughput between GPU workers, with oversubscription based on real collective traffic. It is not the rule that supports the proposed change and acceptance test: A travel assistant has just enough replicas at nominal peak and a single host failure triggers cascading OOM kills. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
Question 216
A data-center AI architecture committee considers resilient interactive service capacity in its AI Infrastructure Components and Architecture review. At the affected site: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. 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.
- Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
- A low-average-latency network can still fail serving SLOs when transient queuing creates high p99 latency.
- AI servers, firmware, NIC drivers, GPU stack and fabric software form a versioned support matrix.
- Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
Explanation
Correct Answer
D:A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
Incorrect Answers
A:Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies. It does not govern the condition described: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
B:A low-average-latency network can still fail serving SLOs when transient queuing creates high p99 latency. It does not govern the condition described: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
C:AI servers, firmware, NIC drivers, GPU stack and fabric software form a versioned support matrix. It does not govern the condition described: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
Question 217
An infrastructure diagnostician is investigating resilient interactive service capacity for the AI Infrastructure Components and Architecture team. The reported incident reads: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. 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.
- Review rack nameplate ratings, measured peak draw, dual-feed loading, breaker headroom and PSU redundancy.
- Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
- Record adapter firmware, host kernel, accelerator driver, CUDA/runtime level, Cisco UCS profiles and network release compatibility.
- Inspect topology failure domains, ECMP next hops, routing timers and failover packet-loss measurements.
Explanation
Correct Answer
B:A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
Incorrect Answers
A:Review rack nameplate ratings, measured peak draw, dual-feed loading, breaker headroom and PSU redundancy. That check would not establish the decisive facts for this condition: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
C:Record adapter firmware, host kernel, accelerator driver, CUDA/runtime level, Cisco UCS profiles and network release compatibility. That check would not establish the decisive facts for this condition: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
D:Inspect topology failure domains, ECMP next hops, routing timers and failover packet-loss measurements. That check would not establish the decisive facts for this condition: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
Question 218
During a bounded AI-platform change-control review concerning resilient interactive service capacity, the site owner reports: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. 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.
- Build physically independent eligible paths and test controlled link or leaf failure under active AI traffic.
- Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
- Use an approved durable checkpoint repository with atomic or validated writes and restore tests.
- Benchmark representative dataset shards and tune supported striping, caching or storage tiers based on bottlenecks.
Explanation
Correct Answer
B:A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
Incorrect Answers
A:Build physically independent eligible paths and test controlled link or leaf failure under active AI traffic. That change would leave the reported problem unresolved: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
C:Use an approved durable checkpoint repository with atomic or validated writes and restore tests. That change would leave the reported problem unresolved: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
D:Benchmark representative dataset shards and tune supported striping, caching or storage tiers based on bottlenecks. That change would leave the reported problem unresolved: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
Question 219
Two independent technical reviewers examine a AI Infrastructure Components and Architecture problem involving resilient interactive service capacity. Their shared incident log states: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. 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.
- Count metadata operations, open latency, namespace contention and worker-level file listing times.
- Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
- Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
- Adopt only vendor-supported accelerator-sharing modes and reserve dedicated resources for jobs that need full-device behavior.
- Record adapter firmware, host kernel, accelerator driver, CUDA/runtime level, Cisco UCS profiles and network release compatibility.
Explanation
Correct Answer
B:A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
C:A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
Incorrect Answers
A:Count metadata operations, open latency, namespace contention and worker-level file listing times. It would not provide the evidence needed for the paired diagnosis and follow-up: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
D:Adopt only vendor-supported accelerator-sharing modes and reserve dedicated resources for jobs that need full-device behavior. It is not the scoped response supported by the accompanying evidence: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
E:Record adapter firmware, host kernel, accelerator driver, CUDA/runtime level, Cisco UCS profiles and network release compatibility. It would not provide the evidence needed for the paired diagnosis and follow-up: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
Question 220
The acceptance board must decide whether an AI-infrastructure response to resilient interactive service capacity is fit for controlled operation. The fictional operating facts are: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. 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.
- Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.
- Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
- Coordinate facility-validated cooling distribution and thermal controls before increasing sustained GPU power.
- Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
- A controlled replica-loss drill retains the peak request target and bounded p99 latency.
- Measured p99 inference response remains within budget during the approved peak-burst workload.
Explanation
Correct Answer
B:A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
D:A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
E:A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. A controlled replica-loss drill retains the peak request target and bounded p99 latency. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
Incorrect Answers
A:Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies. It is not the rule that supports the proposed change and acceptance test: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.
C:Coordinate facility-validated cooling distribution and thermal controls before increasing sustained GPU power. It does not implement the governing rule for this approval: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover. A controlled replica-loss drill retains the peak request target and bounded p99 latency.
F:Measured p99 inference response remains within budget during the approved peak-burst workload. It would not independently demonstrate that the control works: A bank's inference gateway scales only after load rises, but model warm-up exceeds customer timeout during spikes. A controlled replica-loss drill retains the peak request target and bounded p99 latency. Reserve and prewarm sufficient inference replicas across independent hosts to meet stated SLO during failover.