cert
cert-1
cert-2

Pass Cisco 300-640 Exam in First Attempt Guaranteed!

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

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

300-640 Premium File

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

Whats Included:

  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates
 
$98.99
$89.99
accept 10 downloads in the last 7 days
block-screenshots
Cisco 300-640 Practice Test 18

Question 170

The acceptance board must decide whether an AI-infrastructure response to gpu memory and model fit is fit for controlled operation. The fictional operating facts are: An inference service selects a small-memory accelerator that loads the model but crashes at concurrent long-context requests. Signoff cannot rely on a good-looking utilization graph, a change ticket alone or an unmeasured promise. The response must include an appropriate governing engineering principle, a scoped intervention and a separate measurable acceptance criterion. Choose EXACTLY THREE statements that meet those roles.

  1. Measured p99 inference response remains within budget during the approved peak-burst workload.
  2. Build physically independent eligible paths and test controlled link or leaf failure under active AI traffic.
  3. Training storage must sustain aggregate sequential and random reads across workers, not just advertised peak bandwidth.
  4. Training or serving completes a representative workload with safe high-water memory headroom.
  5. Revise supported precision, checkpointing or model sharding and select GPU memory capacity based on peak use.
  6. GPU selection must consider model parameters, activation memory, optimizer states, precision and parallelism rather than nominal FLOPS.

Explanation

Correct Answer

D:

An inference service selects a small-memory accelerator that loads the model but crashes at concurrent long-context requests. Training or serving completes a representative workload with safe high-water memory headroom. Revise supported precision, checkpointing or model sharding and select GPU memory capacity based on peak use.

E:

An inference service selects a small-memory accelerator that loads the model but crashes at concurrent long-context requests. Revise supported precision, checkpointing or model sharding and select GPU memory capacity based on peak use. GPU selection must consider model parameters, activation memory, optimizer states, precision and parallelism rather than nominal FLOPS.

F:

An inference service selects a small-memory accelerator that loads the model but crashes at concurrent long-context requests. GPU selection must consider model parameters, activation memory, optimizer states, precision and parallelism rather than nominal FLOPS. Revise supported precision, checkpointing or model sharding and select GPU memory capacity based on peak use.

Incorrect Answers

A:

Measured p99 inference response remains within budget during the approved peak-burst workload. It would not independently demonstrate that the control works: An inference service selects a small-memory accelerator that loads the model but crashes at concurrent long-context requests. Training or serving completes a representative workload with safe high-water memory headroom. Revise supported precision, checkpointing or model sharding and select GPU memory capacity based on peak use.

B:

Build physically independent eligible paths and test controlled link or leaf failure under active AI traffic. It does not implement the governing rule for this approval: An inference service selects a small-memory accelerator that loads the model but crashes at concurrent long-context requests. Revise supported precision, checkpointing or model sharding and select GPU memory capacity based on peak use. Training or serving completes a representative workload with safe high-water memory headroom.

C:

Training storage must sustain aggregate sequential and random reads across workers, not just advertised peak bandwidth. It is not the rule that supports the proposed change and acceptance test: An inference service selects a small-memory accelerator that loads the model but crashes at concurrent long-context requests. GPU selection must consider model parameters, activation memory, optimizer states, precision and parallelism rather than nominal FLOPS. Revise supported precision, checkpointing or model sharding and select GPU memory capacity based on peak use.

Question 171

A data-center AI architecture committee considers numa locality and data-loader cpu in its AI Infrastructure Components and Architecture review. At the affected site: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. The committee is deliberately not selecting a vendor product or applying a live change yet. It needs the foundational reasoning that makes the next design decision defensible. Which ONE engineering principle directly governs this situation? Choose ONE answer.

  1. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.
  2. Serving capacity needs headroom for node failure, peak concurrency and warm model loading time.
  3. Storage capacity plans require dataset lineage, scratch amplification, checkpoint retention and growth forecasts, not initial dataset size only.
  4. GPU partitioning and virtualization require platform-licensed, driver-compatible capabilities verified for the target workload.

Explanation

Correct Answer

A:

An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.

Incorrect Answers

B:

Serving capacity needs headroom for node failure, peak concurrency and warm model loading time. It does not govern the condition described: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.

C:

Storage capacity plans require dataset lineage, scratch amplification, checkpoint retention and growth forecasts, not initial dataset size only. It does not govern the condition described: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.

D:

GPU partitioning and virtualization require platform-licensed, driver-compatible capabilities verified for the target workload. It does not govern the condition described: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.

Question 172

An infrastructure diagnostician is investigating numa locality and data-loader cpu for the AI Infrastructure Components and Architecture team. The reported incident reads: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Competing teams attribute the symptoms to different layers of the AI stack. Before accepting a hypothesis or touching production settings, the investigator must choose a discriminating set of measurements. Which ONE evidence package is MOST relevant? Choose ONE answer.

  1. Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity.
  2. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.
  3. Collect out-of-memory traces, per-rank allocator usage, peak activations, tensor sizes and chosen precision.
  4. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.

Explanation

Correct Answer

B:

An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.

Incorrect Answers

A:

Model concurrent requests, cold-start time, per-replica memory, autoscaling delay and one-node-failed capacity. That check would not establish the decisive facts for this condition: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.

C:

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: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.

D:

Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements. That check would not establish the decisive facts for this condition: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.

Question 173

During a bounded AI-platform change-control review concerning numa locality and data-loader cpu, the site owner reports: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. The proposed work must have a narrow implementation scope, first be validated against the failure evidence and avoid unrelated system-wide tuning. Which ONE intervention should the engineers pilot before any wider rollout? Choose ONE answer.

  1. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.
  2. Build physically independent eligible paths and test controlled link or leaf failure under active AI traffic.
  3. Size and balance the leaf-spine fabric for the observed communication pattern before adding accelerator nodes.
  4. Coordinate facility-validated cooling distribution and thermal controls before increasing sustained GPU power.

Explanation

Correct Answer

A:

An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.

Incorrect Answers

B:

Build physically independent eligible paths and test controlled link or leaf failure under active AI traffic. That change would leave the reported problem unresolved: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

C:

Size and balance the leaf-spine fabric for the observed communication pattern before adding accelerator nodes. That change would leave the reported problem unresolved: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

D:

Coordinate facility-validated cooling distribution and thermal controls before increasing sustained GPU power. That change would leave the reported problem unresolved: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

Question 174

Two independent technical reviewers examine a AI Infrastructure Components and Architecture problem involving numa locality and data-loader cpu. Their shared incident log states: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. One reviewer requires measurements capable of separating the observed problem from neighboring failure modes; the other requires a bounded action that follows those measurements. Choose EXACTLY TWO statements that together supply those two parts.

  1. Set class-based retention and quotas while keeping legally required provenance and restorable checkpoints.
  2. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.
  3. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.
  4. Correlate request p99 with per-hop queue occupancy, drops, ECN marks and burst timing rather than average link load.
  5. Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements.

Explanation

Correct Answer

B:

An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. GPU duty cycle and samples-per-second improve with no host memory thrashing.

C:

An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality. GPU duty cycle and samples-per-second improve with no host memory thrashing.

Incorrect Answers

A:

Set class-based retention and quotas while keeping legally required provenance and restorable checkpoints. It is not the scoped response supported by the accompanying evidence: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. GPU duty cycle and samples-per-second improve with no host memory thrashing.

D:

Correlate request p99 with per-hop queue occupancy, drops, ECN marks and burst timing rather than average link load. It would not provide the evidence needed for the paired diagnosis and follow-up: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality. GPU duty cycle and samples-per-second improve with no host memory thrashing.

E:

Map node-to-rack, PDU, network-fabric and storage failure domains alongside worker-placement requirements. It would not provide the evidence needed for the paired diagnosis and follow-up: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality. GPU duty cycle and samples-per-second improve with no host memory thrashing.

Question 175

The acceptance board must decide whether an AI-infrastructure response to numa locality and data-loader cpu is fit for controlled operation. The fictional operating facts are: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Signoff cannot rely on a good-looking utilization graph, a change ticket alone or an unmeasured promise. The response must include an appropriate governing engineering principle, a scoped intervention and a separate measurable acceptance criterion. Choose EXACTLY THREE statements that meet those roles.

  1. Use an approved durable checkpoint repository with atomic or validated writes and restore tests.
  2. Training or serving completes a representative workload with safe high-water memory headroom.
  3. GPU duty cycle and samples-per-second improve with no host memory thrashing.
  4. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.
  5. GPU partitioning and virtualization require platform-licensed, driver-compatible capabilities verified for the target workload.
  6. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.

Explanation

Correct Answer

C:

An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. GPU duty cycle and samples-per-second improve with no host memory thrashing. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

D:

An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.

F:

An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

Incorrect Answers

A:

Use an approved durable checkpoint repository with atomic or validated writes and restore tests. It does not implement the governing rule for this approval: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. GPU duty cycle and samples-per-second improve with no host memory thrashing.

B:

Training or serving completes a representative workload with safe high-water memory headroom. It would not independently demonstrate that the control works: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. GPU duty cycle and samples-per-second improve with no host memory thrashing. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

E:

GPU partitioning and virtualization require platform-licensed, driver-compatible capabilities verified for the target workload. It is not the rule that supports the proposed change and acceptance test: An autonomous-vehicle trainer shows only 40 percent GPU utilization while CPU image decoding maxes one NUMA node. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

Question 176

A data-center AI architecture committee considers numa locality and data-loader cpu in its AI Infrastructure Components and Architecture review. At the affected site: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. The committee is deliberately not selecting a vendor product or applying a live change yet. It needs the foundational reasoning that makes the next design decision defensible. Which ONE engineering principle directly governs this situation? Choose ONE answer.

  1. Checkpoint systems must meet recovery-point and recovery-time needs with verified durable writes and restore validation.
  2. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.
  3. A low-average-latency network can still fail serving SLOs when transient queuing creates high p99 latency.
  4. Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies.

Explanation

Correct Answer

B:

A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.

Incorrect Answers

A:

Checkpoint systems must meet recovery-point and recovery-time needs with verified durable writes and restore validation. It does not govern the condition described: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.

C:

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 genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.

D:

Distributed AI placement must limit correlated failures and balance locality against rack, host and power dependencies. It does not govern the condition described: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.

Question 177

An infrastructure diagnostician is investigating numa locality and data-loader cpu for the AI Infrastructure Components and Architecture team. The reported incident reads: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Competing teams attribute the symptoms to different layers of the AI stack. Before accepting a hypothesis or touching production settings, the investigator must choose a discriminating set of measurements. Which ONE evidence package is MOST relevant? Choose ONE answer.

  1. Review rack nameplate ratings, measured peak draw, dual-feed loading, breaker headroom and PSU redundancy.
  2. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.
  3. Compare inlet temperatures, coolant flow where relevant, facility heat rejection, exhaust recirculation and throttling events.
  4. Record adapter firmware, host kernel, accelerator driver, CUDA/runtime level, Cisco UCS profiles and network release compatibility.

Explanation

Correct Answer

B:

A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.

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 genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.

C:

Compare inlet temperatures, coolant flow where relevant, facility heat rejection, exhaust recirculation and throttling events. That check would not establish the decisive facts for this condition: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.

D:

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 genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.

Question 178

During a bounded AI-platform change-control review concerning numa locality and data-loader cpu, the site owner reports: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. The proposed work must have a narrow implementation scope, first be validated against the failure evidence and avoid unrelated system-wide tuning. Which ONE intervention should the engineers pilot before any wider rollout? Choose ONE answer.

  1. Benchmark representative dataset shards and tune supported striping, caching or storage tiers based on bottlenecks.
  2. Reserve appropriate service classes and relieve burst contention after validating which path contributes to tail delay.
  3. Compare like-for-like PUE baselines and optimize cooling or capacity with a documented energy measurement plan.
  4. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

Explanation

Correct Answer

D:

A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.

Incorrect Answers

A:

Benchmark representative dataset shards and tune supported striping, caching or storage tiers based on bottlenecks. That change would leave the reported problem unresolved: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

B:

Reserve appropriate service classes and relieve burst contention after validating which path contributes to tail delay. That change would leave the reported problem unresolved: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

C:

Compare like-for-like PUE baselines and optimize cooling or capacity with a documented energy measurement plan. That change would leave the reported problem unresolved: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

Question 179

Two independent technical reviewers examine a AI Infrastructure Components and Architecture problem involving numa locality and data-loader cpu. Their shared incident log states: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. One reviewer requires measurements capable of separating the observed problem from neighboring failure modes; the other requires a bounded action that follows those measurements. Choose EXACTLY TWO statements that together supply those two parts.

  1. Package or shard files appropriately and scale the supported metadata path while preserving data semantics.
  2. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.
  3. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality.
  4. Count metadata operations, open latency, namespace contention and worker-level file listing times.
  5. Inspect topology failure domains, ECMP next hops, routing timers and failover packet-loss measurements.

Explanation

Correct Answer

B:

A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. GPU duty cycle and samples-per-second improve with no host memory thrashing.

C:

A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality. GPU duty cycle and samples-per-second improve with no host memory thrashing.

Incorrect Answers

A:

Package or shard files appropriately and scale the supported metadata path while preserving data semantics. It is not the scoped response supported by the accompanying evidence: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. GPU duty cycle and samples-per-second improve with no host memory thrashing.

D:

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 genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality. GPU duty cycle and samples-per-second improve with no host memory thrashing.

E:

Inspect topology failure domains, ECMP next hops, routing timers and failover packet-loss measurements. It would not provide the evidence needed for the paired diagnosis and follow-up: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Observe GPU idle gaps, CPU saturation, page faults, dataloader queue depth, NUMA affinity and PCIe locality. GPU duty cycle and samples-per-second improve with no host memory thrashing.

Question 180

The acceptance board must decide whether an AI-infrastructure response to numa locality and data-loader cpu is fit for controlled operation. The fictional operating facts are: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Signoff cannot rely on a good-looking utilization graph, a change ticket alone or an unmeasured promise. The response must include an appropriate governing engineering principle, a scoped intervention and a separate measurable acceptance criterion. Choose EXACTLY THREE statements that meet those roles.

  1. The same-load measurement demonstrates reduced facility overhead without harming service performance.
  2. AI thermal designs must handle rack heat flux and liquid or air-cooling capacity under worst-case workloads.
  3. GPU duty cycle and samples-per-second improve with no host memory thrashing.
  4. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.
  5. Adopt only vendor-supported accelerator-sharing modes and reserve dedicated resources for jobs that need full-device behavior.
  6. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

Explanation

Correct Answer

C:

A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. GPU duty cycle and samples-per-second improve with no host memory thrashing. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

D:

A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

F:

A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation.

Incorrect Answers

A:

The same-load measurement demonstrates reduced facility overhead without harming service performance. It would not independently demonstrate that the control works: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. GPU duty cycle and samples-per-second improve with no host memory thrashing. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

B:

AI thermal designs must handle rack heat flux and liquid or air-cooling capacity under worst-case workloads. It is not the rule that supports the proposed change and acceptance test: A genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. A GPU pipeline needs adequate CPU, RAM and local device bandwidth to feed accelerators without host-side starvation. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs.

E:

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 genomics workload adds accelerators but each device waits on a host preprocessing thread pinned across sockets. Tune CPU worker allocation, memory pinning and NUMA or device placement before buying additional GPUs. GPU duty cycle and samples-per-second improve with no host memory thrashing.

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