I85PREP
⌘ K
300-640 DCAI study materials
300-640 DCAI study resources

300-640 DCAI practice questions

One question at a time. Build your 300-640 DCAI understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1AI Infrastructure Architecture and Design

In a Cisco AI infrastructure design, which fabric topology is typically recommended for a large-scale GPU training cluster that requires predictable any-to-any latency and full bisectional bandwidth?

  1. A three-tier core-aggregation-access campus design
  2. A two-tier leaf-spine fabric with equal-cost multipathing Correct answer.
  3. A ring topology connecting all GPU servers in sequence
  4. A single-tier hub-and-spoke topology with a central aggregation switch

Explanation

Large GPU training clusters need non-blocking, any-to-any communication so that collective operations like all-reduce are not bottlenecked. A two-tier leaf-spine fabric provides consistent hop count and ECMP load sharing, which delivers predictable latency and full bisectional bandwidth.

Question 2AI Infrastructure Architecture and Design

Which Ethernet technology allows lossless transport of RDMA traffic over a converged Cisco data center network?

  1. Priority Flow Control with Data Center Bridging Correct answer.
  2. Spanning Tree Protocol root guard
  3. Standard 802.1Q trunking with native VLAN
  4. Link Aggregation Control Protocol bundling

Explanation

RDMA over Converged Ethernet requires a lossless class of service so that dropped frames do not trigger retransmissions that degrade GPU communication. Priority Flow Control, part of Data Center Bridging, pauses only the affected traffic class to achieve this losslessness.

Question 3AI Infrastructure Architecture and Design

A design team must choose a storage access model for a training pipeline where many GPU workers read the same large dataset concurrently. Which approach best supports high aggregate throughput with shared caching?

  1. USB-attached external drives per node
  2. A single NFS server with one 10 Gigabit uplink
  3. A parallel scale-out file system accessed over a high-speed fabric Correct answer.
  4. Direct-attached SAS disks on each GPU server

Explanation

Concurrent reads by many workers demand aggregate bandwidth that scales with the number of storage nodes. A parallel scale-out file system distributes data and metadata across nodes and supports shared caching, avoiding a single-server bottleneck.

Question 4AI Infrastructure Architecture and Design

Which Cisco platform provides a centralized policy and automation controller for intent-based networking across data center and AI fabrics?

  1. Cisco Identity Services Engine
  2. Cisco Catalyst Center
  3. Cisco Nexus Dashboard Correct answer.
  4. Cisco Umbrella

Explanation

Nexus Dashboard is the operations and policy hub for Cisco data center networking, including fabric controllers and assurance for AI-ready fabrics. It aggregates fabric management, telemetry, and orchestration in one place.

Question 5AI Infrastructure Architecture and Design

What is the primary reason to design separate front-end and back-end networks in a GPU cluster?

  1. To reduce the number of required IP addresses
  2. To allow the cluster to operate without any switching hardware
  3. To isolate bulk east-west training traffic from management and storage flows Correct answer.
  4. To eliminate the need for routing protocols

Explanation

GPU collective traffic is bursty and latency-sensitive, so separating it from storage and management flows prevents contention. Dedicated back-end fabrics let each network be sized and tuned for its own traffic profile.

Question 6AI Infrastructure Architecture and Design

Which power and cooling consideration most directly constrains the maximum number of high-density GPU servers that can be placed in a single rack?

  1. The number of available switch ports
  2. The server operating system version
  3. Per-rack power budget and airflow capacity Correct answer.
  4. The length of fiber optic patch cables

Explanation

GPU servers draw very high wattage, so the rack's electrical feed and cooling airflow set the practical density ceiling. Designers must balance compute density against per-rack power and thermal limits.

Question 7AI Infrastructure Architecture and Design

In an AI cluster design, what does the term oversubscription ratio describe?

  1. The ratio of GPU memory to system memory
  2. The percentage of storage capacity reserved for snapshots
  3. The ratio of aggregate downlink bandwidth to uplink bandwidth at a tier Correct answer.
  4. The number of virtual machines per physical host

Explanation

Oversubscription expresses how much more bandwidth is available toward endpoints than toward the upper tier. A 1:1 ratio means non-blocking, while higher ratios reduce cost at the expense of contention during collective traffic.

Question 8AI Infrastructure Architecture and Design

Which capability of Cisco Nexus switches supports telemetry streaming that helps operators detect microbursts in AI fabrics?

  1. Model-driven streaming telemetry with high-frequency counters Correct answer.
  2. Port mirroring to a single management port
  3. Static ARP entries
  4. Spanning tree topology change notifications

Explanation

Microbursts are short congestion spikes that average-rate polling misses, so high-frequency streaming telemetry is needed. Model-driven telemetry pushes fine-grained counters to collectors for near-real-time analysis.

Question 9AI Infrastructure Architecture and Design

An architect is sizing a leaf-spine fabric for 512 GPU servers, each with two 400G back-end ports. Which design decision most directly preserves non-blocking behavior?

  1. Providing spine uplink capacity equal to the sum of leaf downlink capacity Correct answer.
  2. Reducing the number of spines to lower cost
  3. Using a single leaf switch for all servers
  4. Connecting all servers to one spine

Explanation

Non-blocking behavior requires that uplink bandwidth at each tier match the aggregate downlink bandwidth. With 512 dual-400G servers, the spine tier must supply enough capacity to absorb the full leaf downlink total.

Question 10AI Infrastructure Architecture and Design

During an all-reduce operation across 256 GPUs, operators observe periodic throughput dips correlated with congestion on a specific spine link. Which design change best addresses this?

  1. Reduce the number of leaf switches
  2. Move all training traffic to the management network
  3. Disable ECMP on the fabric
  4. Increase path diversity with more equal-cost spine links and tune load balancing Correct answer.

Explanation

Collective traffic patterns can create hot spots when hashing is uneven, so more equal-cost paths plus tuned load balancing spread flows better. This reduces congestion on any single spine link.

Question 11AI Infrastructure Architecture and Design

A customer wants to run both training and inference workloads on the same fabric while guaranteeing latency for inference. Which design approach best meets this requirement?

  1. Route inference traffic through the internet edge
  2. Disable flow control on all ports
  3. Combine all traffic into a single best-effort class
  4. Implement QoS classes that prioritize latency-sensitive inference traffic over bulk training flows Correct answer.

Explanation

Mixing workloads requires differentiated treatment so that bursty training traffic cannot starve latency-sensitive inference. QoS classification with prioritized queuing and flow control preserves inference latency targets.

Question 12AI Infrastructure Architecture and Design

An architect must provide tenant isolation for multiple research teams sharing a GPU fabric. Which combination best achieves isolation while retaining shared infrastructure efficiency?

  1. Separate physical fabrics for every team
  2. A single flat VLAN for all teams
  3. Per-team spanning tree instances
  4. VRF segmentation with per-tenant QoS policies on a shared physical fabric Correct answer.

Explanation

VRFs provide logical separation of routing and addressing on shared hardware, and per-tenant QoS keeps one team's traffic from degrading another's. This preserves efficiency while enforcing isolation.

Question 13AI Infrastructure Architecture and Design

A design requires that a single failed spine switch not reduce aggregate fabric bandwidth below the non-blocking threshold. Which approach satisfies this?

  1. Rely on a single high-capacity spine
  2. Reduce leaf uplinks when a spine fails
  3. Accept oversubscription only during failures
  4. Size the spine tier with N+1 redundancy so remaining spines carry the full load Correct answer.

Explanation

N+1 spine redundancy ensures that if one spine fails, the remaining spines still provide enough uplink capacity to stay non-blocking. Capacity must be provisioned for the failure case, not just steady state.

Question 14AI Infrastructure Architecture and Design

Which storage design pattern best supports checkpointing from a large distributed training job without stalling GPU computation?

  1. Storing checkpoints only in GPU memory
  2. A high-throughput parallel file system with sufficient write bandwidth and tiering Correct answer.
  3. A single low-cost SATA drive shared over NFS
  4. Writing checkpoints to local USB storage

Explanation

Checkpoint writes are large and frequent, so the storage tier must absorb them without back-pressuring the training job. A parallel file system with adequate write bandwidth and tiering prevents checkpoint I/O from becoming the bottleneck.

Question 15AI Infrastructure Architecture and Design

An architect is evaluating RoCEv2 for a GPU back-end fabric. Which configuration is most important to prevent packet drops that degrade RDMA performance?

  1. Increasing the MTU to 9000 bytes only
  2. Using untagged frames on all links
  3. Disabling all congestion notification
  4. End-to-end lossless class of service with PFC and ECN marking Correct answer.

Explanation

RDMA assumes a lossless fabric, so PFC prevents drops at congestion points while ECN signals senders to slow down before buffers overflow. Together they maintain the low-latency behavior RDMA requires.

Question 16AI Infrastructure Architecture and Design

A customer needs an AI fabric that supports both 400G and 100G server connections with a consistent management model. Which design choice best meets this?

  1. Direct server-to-server cabling without switches
  2. A single fixed switch model for all roles
  3. A leaf-spine fabric using modular and fixed Nexus switches managed through Nexus Dashboard Correct answer.
  4. A hub-and-spoke design with mixed vendor management tools

Explanation

A leaf-spine fabric with a mix of modular and fixed Nexus platforms accommodates multiple port speeds, and Nexus Dashboard provides one consistent management and policy model. This balances flexibility with operational simplicity.

Question 17AI Infrastructure Architecture and Design

Which factor most influences the choice between air cooling and liquid cooling for a high-density GPU cluster?

  1. Per-rack thermal design power and the ability of airflow to remove heat Correct answer.
  2. The number of VLANs configured on the fabric
  3. The routing protocol selected for the underlay
  4. The file system used for checkpoints

Explanation

As rack power density rises, air cooling reaches practical limits because airflow cannot remove heat fast enough. Liquid cooling becomes necessary when per-rack thermal design power exceeds what air handling can support.

Question 18AI Infrastructure Architecture and Design

An architect is designing an inference platform that must scale horizontally and serve many small requests with low latency. Which design element is most important?

  1. Routing all requests through a central database
  2. A single monolithic inference server
  3. Load balancing across many inference nodes with health checking Correct answer.
  4. Batch-only processing with long queues

Explanation

Low-latency inference at scale requires distributing requests across many healthy nodes so no single server becomes a bottleneck. Load balancing with health checks maintains availability and responsiveness.

Question 19AI Infrastructure Architecture and Design

A design team observes that GPU utilization drops whenever the storage network shares physical links with the training fabric. Which remediation is most appropriate?

  1. Separate the storage and training traffic onto dedicated fabric paths or classes Correct answer.
  2. Disable telemetry to reduce overhead
  3. Increase the number of spines without changing traffic separation
  4. Reduce GPU count to lower demand

Explanation

Contention on shared links causes storage traffic to compete with latency-sensitive training flows, reducing GPU utilization. Dedicated paths or strict class separation prevents storage I/O from interfering with collectives.

Question 20AI Infrastructure Architecture and Design

An architect must justify a non-blocking fabric to a cost-conscious stakeholder. Which argument best explains the trade-off?

  1. Oversubscription improves GPU utilization during all-reduce
  2. Non-blocking fabrics always cost less than oversubscribed fabrics
  3. Fabric bandwidth has no effect on training time
  4. Oversubscription saves capital cost but lengthens collective completion times and idles expensive GPUs Correct answer.

Explanation

GPU compute is far more expensive than fabric ports, so idle GPUs waiting on congested links can cost more than the savings from oversubscription. Non-blocking design protects the return on the most costly resource.

Question 21AI Infrastructure Architecture and Design

An operator reports that a single leaf switch failure caused a training job to fail rather than degrade. Which design weakness is most likely responsible?

  1. The fabric used too many equal-cost paths
  2. Servers were dual-homed but the job lacked resilient collective communication and failover configuration Correct answer.
  3. The spine tier was over-provisioned
  4. QoS was configured with multiple classes

Explanation

Physical dual-homing alone does not guarantee job survival; the collective library and job scheduler must also detect and recover from a failed path. Without resilient communication configuration, a single leaf failure aborts the job.

Question 22AI Infrastructure Architecture and Design

A design must support both RoCEv2 lossless traffic and standard TCP traffic on the same leaf ports. Which configuration approach is most appropriate?

  1. Apply PFC to all traffic classes on the port
  2. Mark all traffic as RoCEv2 to simplify configuration
  3. Disable QoS and rely on best-effort delivery
  4. Classify RoCEv2 into a no-drop class with PFC while leaving TCP in a drop-enabled class Correct answer.

Explanation

PFC should be scoped to the lossless class so that pause frames do not stall unrelated TCP traffic, which could create head-of-line blocking. Classifying only RoCEv2 as no-drop preserves RDMA performance without harming other flows.

Question 23AI Infrastructure Architecture and Design

An architect is comparing a single large fabric against multiple smaller pods for a multi-thousand-GPU deployment. Which consideration most favors a pod-based design?

  1. Guaranteeing zero oversubscription across all pods
  2. Eliminating all inter-pod traffic
  3. Containing failure domains and traffic locality while scaling through spine aggregation Correct answer.
  4. Removing the need for routing protocols

Explanation

Pods limit the blast radius of failures and keep most collective traffic local, while spine aggregation connects pods for cross-pod communication. This scales the fabric without creating one enormous failure domain.

Question 24AI Infrastructure Architecture and Design

During capacity planning, an architect calculates that peak collective traffic briefly exceeds average fabric capacity. Which design principle best addresses this burst behavior?

  1. Reduce telemetry sampling to hide the peaks
  2. Size the fabric strictly to average utilization
  3. Rely on retransmissions to recover from drops
  4. Provision headroom and use buffering with congestion control to absorb transient peaks Correct answer.

Explanation

Collective operations produce synchronized bursts that can exceed average demand, so headroom and buffering with congestion control absorb these transients. Sizing only to averages leaves no margin for burst peaks.

Question 25AI Infrastructure Architecture and Design

An architect must choose an underlay routing approach for a large AI fabric that requires fast convergence and simple operations. Which choice is most appropriate?

  1. A single default route to the core
  2. Static routes on every switch
  3. A link-state interior gateway protocol such as OSPF or IS-IS with ECMP Correct answer.
  4. RIP version 1

Explanation

A link-state IGP provides fast convergence and supports equal-cost multipathing, which is essential for spreading flows across a leaf-spine fabric. Static or default-only routing cannot adapt quickly to failures at scale.

Question 26AI Infrastructure Architecture and Design

A GPU server vendor specifies a thermal design power of 10 kW per rack for a new training cluster. Which infrastructure decision is most directly driven by this figure?

  1. Choosing the routing protocol used on the back-end GPU fabric
  2. Determining the RAID level for the training dataset volume
  3. Setting the maximum transmission unit on the front-end management network
  4. Selecting a cooling distribution approach capable of removing 10 kW from the rack envelope Correct answer.

Explanation

Rack power density determines how heat is removed, so it drives the choice between air cooling, rear-door heat exchangers, and direct-to-chip liquid cooling. Network protocol, storage RAID level, and MTU settings are unrelated to the thermal envelope of the rack.

Question 27AI Infrastructure Architecture and Design

An architect is documenting the management plane for a GPU cluster. Which element belongs to the out-of-band management design rather than the data plane?

  1. The overlay tunnel endpoints used by tenant workloads
  2. The leaf-spine links carrying training gradients
  3. Baseboard management controller access for firmware and console recovery Correct answer.
  4. Remote direct memory access queues used by collective operations

Explanation

Out-of-band management uses the baseboard management controller on a separate network so operators can reach a server even when the host OS or data plane is down. RDMA queues, fabric links, and overlay endpoints all carry or support production workload traffic.

Question 28AI Infrastructure Architecture and Design

A design team must place GPU servers so that a single top-of-rack failure affects the fewest collective participants. Which placement principle supports that goal?

  1. Assign each tenant a dedicated spine switch to isolate failures
  2. Connect every GPU directly to a spine to remove leaf dependency
  3. Distribute each job's ranks across many racks while keeping rail-aligned cabling consistent Correct answer.
  4. Concentrate all ranks of one job in a single rack to shorten cable runs

Explanation

Spreading ranks across racks limits the blast radius of any single top-of-rack failure, while consistent rail-aligned cabling preserves efficient collective paths. Concentrating a job in one rack maximizes the impact of that rack's failure.

Question 29AI Infrastructure Architecture and Design

Which statement correctly distinguishes the front-end network from the back-end network in a typical AI training cluster design?

  1. The front-end must be non-blocking, while the back-end may be heavily oversubscribed
  2. The back-end carries GPU-to-GPU collective traffic, while the front-end carries storage, user, and orchestration traffic Correct answer.
  3. Both networks must share the same physical switches to guarantee equal latency
  4. The back-end carries user login sessions, while the front-end carries collective traffic

Explanation

The back-end fabric is purpose-built for east-west GPU communication such as all-reduce, and the front-end handles north-south and storage or orchestration flows. Keeping them distinct prevents storage and user traffic from perturbing latency-sensitive collectives.

Question 30AI Infrastructure Architecture and Design

An architect must choose between a two-tier leaf-spine fabric and a three-tier fabric with a super-spine layer for a planned 8,000-GPU cluster. Which consideration most strongly favors the three-tier design?

  1. The cluster needs a single management IP address for all switches
  2. The workload requires jumbo frames on the storage network
  3. The number of leaf and spine ports available cannot provide enough non-blocking paths at the required radix Correct answer.
  4. The customer wants to reuse existing 1G switches for management

Explanation

A super-spine tier is introduced when the port radix of leaf and spine switches cannot support the required number of endpoints at the target bandwidth without oversubscription. Management addressing, jumbo frames, and legacy management switches do not drive tier count.

Question 31AI Infrastructure Architecture and Design

A design specifies a 1:1 subscription ratio on the GPU back-end fabric. Which outcome does this ratio describe?

  1. Each server is guaranteed exactly one terabit of bandwidth
  2. Each leaf connects to exactly one spine for redundancy
  3. The fabric reserves half of its capacity for storage traffic
  4. Every server-facing port can reach the core at full line rate simultaneously Correct answer.

Explanation

A 1:1 subscription, or non-blocking ratio, means the aggregate uplink capacity from the access layer matches the aggregate downlink capacity to endpoints, so all servers can transmit at line rate at once. It is a capacity ratio, not a bandwidth guarantee in terabits or a reservation for storage.

Question 32AI Infrastructure Architecture and Design

An operator needs to confirm that the GPU back-end fabric will keep working when one spine fails. Which design property should the architect verify?

  1. Remaining spine capacity still meets or exceeds the required non-blocking threshold for the attached leaves Correct answer.
  2. The failed spine is the designated backup for all storage VLANs
  3. Each leaf has at least one uplink to the failed spine for fast reconvergence
  4. Spanning tree is enabled on all leaf-spine links to block loops

Explanation

N-minus-one resilience means that after losing a spine, the surviving spines still provide the uplink capacity needed to keep the fabric at the intended subscription ratio. Keeping an uplink to a failed spine or relying on spanning tree does not restore lost capacity.

Question 33AI Infrastructure Architecture and Design

A customer wants each research team's traffic to remain isolated on a shared GPU fabric without building separate physical switches per team. Which design approach meets this requirement?

  1. MAC address filtering on each server port to separate teams
  2. VRF instances per team combined with overlay segmentation and per-tenant QoS policies Correct answer.
  3. One dedicated spine per team with all leaves shared across teams
  4. A single flat VLAN shared by all teams with priority queuing for storage

Explanation

VRFs plus overlay segmentation give logical separation of routing and forwarding domains on shared hardware, and per-tenant QoS keeps one team's bursts from harming another. A flat VLAN, MAC filtering, or per-team spines on shared leaves do not provide complete isolation.

Question 34AI Infrastructure Architecture and Design

A storage team proposes running checkpoint writes over the same leaf ports that carry GPU collective traffic. Which design response best protects training throughput?

  1. Provide a dedicated storage fabric or separate uplinks so checkpoint bursts do not contend with collective traffic Correct answer.
  2. Increase the MTU on the shared ports to absorb the checkpoint bursts
  3. Move checkpointing to the out-of-band management network
  4. Enable priority flow control so checkpoints are paused during collectives

Explanation

Checkpoint writes are large bursty flows, and separating them onto dedicated links or a dedicated storage fabric prevents them from consuming buffer and bandwidth needed by latency-sensitive collectives. MTU changes and pause frames do not remove the underlying contention.

Question 35AI Infrastructure Architecture and Design

An architect is selecting a back-end transport for a new GPU fabric and wants congestion control that operates at the endpoint without requiring a lossless hop-by-hop fabric. Which approach fits that requirement?

  1. A congestion-controlled RDMA transport that reacts to explicit congestion notification at the endpoints Correct answer.
  2. A transport that depends on priority flow control across every hop to avoid drops
  3. A transport that relies on spanning tree to prevent loops in the fabric
  4. A transport that disables ECN and relies on TCP slow start only

Explanation

Endpoint-based congestion control that responds to ECN allows the fabric to signal congestion before buffers overflow, reducing dependence on a fully lossless hop-by-hop design. Relying solely on priority flow control or disabling ECN reintroduces head-of-line blocking and drop-synchronization risks.

Question 36AI Infrastructure Architecture and Design

A design must support both latency-sensitive inference traffic and throughput-oriented batch training on the same physical leaf-spine fabric. Which combination best meets that goal?

  1. Strict-priority queuing for inference classes with weighted scheduling for training, plus admission limits on the inference queue Correct answer.
  2. Disabling ECN so that inference and training share the same drop behavior
  3. A single best-effort queue for all traffic with larger buffers on every port
  4. Equal-cost multipath hashing that ignores traffic class

Explanation

Priority treatment for latency-sensitive classes combined with weighted scheduling for bulk traffic protects inference latency while still letting training use spare capacity. Admission limits prevent the priority queue from starving other classes.

Question 37AI Infrastructure Architecture and Design

An architect is comparing a single large non-blocking fabric against several smaller pods interconnected by a higher-tier layer for a multi-thousand-GPU deployment. Which trade-off most accurately describes the pod-based option?

  1. It guarantees lower inter-pod latency than a single flat fabric at the same cost
  2. It removes the need for any spine layer within each pod
  3. It eliminates oversubscription entirely because pods are independent
  4. It scales more granularly and contains failures per pod, but inter-pod collectives may see higher latency and oversubscription Correct answer.

Explanation

Pod-based designs let operators grow in increments and contain faults within a pod, but traffic that crosses pods traverses additional tiers and may face oversubscription. A single flat fabric avoids extra hops but is harder to grow and has a larger failure domain.

Question 38AI Infrastructure Architecture and Design

During peak collective phases, traffic briefly exceeds average fabric capacity but stays below the capacity of the combined uplinks. Which design element most directly absorbs this burst without dropping packets?

  1. Reducing the number of spines to lower path diversity
  2. Increasing the number of MAC addresses learned per leaf
  3. Sufficient buffering and ECN marking thresholds sized for the burst duration and microburst pattern Correct answer.
  4. Disabling equal-cost multipath to keep flows on one path

Explanation

Microburst absorption depends on buffer sizing and ECN thresholds tuned to the burst duration and traffic pattern. Reducing path diversity or MAC learning capacity does not create the ability to hold transient excess traffic.

Question 39AI Infrastructure Architecture and Design

An architect must justify a non-blocking back-end fabric to a stakeholder who prefers a cheaper oversubscribed design. Which argument is most technically sound?

  1. Oversubscribed fabrics cannot carry any traffic during collective operations
  2. Collectives synchronize all ranks, so a congested uplink stalls the entire job rather than only the affected server Correct answer.
  3. Oversubscribed designs require more spine switches than non-blocking designs
  4. Non-blocking fabrics remove the need for congestion control at the endpoints

Explanation

Because collective operations such as all-reduce require every rank to finish before the next step, congestion on any shared uplink delays the slowest participant and idles the whole job. This bulk-synchronous behavior is why oversubscription is especially costly for training.

Question 40AI Infrastructure Architecture and Design

An operator reports that a single leaf switch reboot caused a long-running training job to abort instead of pausing. Which design weakness best explains this behavior?

  1. The fabric used equal-cost multipath across all spines
  2. The endpoints used congestion control based on ECN marking
  3. The storage network was physically separate from the compute fabric
  4. The job's ranks were concentrated on that leaf with no alternative path or checkpoint restart strategy Correct answer.

Explanation

If many ranks of one job sit behind the same leaf and the application has no checkpoint or restart strategy, losing that leaf terminates the collective and the job. Spreading ranks and checkpointing converts a hard failure into a recoverable pause.

Question 41Compute and GPU Fabric for AI Workloads

In a Cisco data center AI fabric, which GPU interconnect technology is used for direct high-bandwidth GPU-to-GPU communication within a single server node?

  1. PCIe Gen3 x16
  2. Ethernet RoCEv2
  3. InfiniBand HDR
  4. NVLink Correct answer.

Explanation

NVLink is a dedicated high-bandwidth GPU-to-GPU interconnect that lets GPUs within a node share data directly without traversing the PCIe bus or the network. InfiniBand and RoCEv2 are used for inter-node communication, not intra-node GPU peering.

Question 42Compute and GPU Fabric for AI Workloads

Which fabric characteristic most directly determines how quickly distributed training workers can exchange gradient tensors across nodes?

  1. Storage IOPS on the shared NAS
  2. North-south bandwidth to the internet edge
  3. Number of CPU cores per node
  4. East-west bandwidth and latency between compute nodes Correct answer.

Explanation

Distributed training generates heavy all-reduce and all-gather traffic between compute nodes, so east-west bandwidth and latency govern how fast gradient exchange completes. North-south paths and storage IOPS affect other flows but not the node-to-node collective performance.

Question 43Compute and GPU Fabric for AI Workloads

What is the primary purpose of a rail-optimized network design in a GPU cluster?

  1. To isolate management traffic from production traffic
  2. To provide redundant power feeds to each rack
  3. To ensure each GPU connects to a separate leaf switch so collectives avoid oversubscription Correct answer.
  4. To compress gradient traffic before transmission

Explanation

Rail-optimized designs map each GPU in a server to its own leaf switch rail, so identical-index GPUs communicate over dedicated non-oversubscribed paths. This reduces contention during collective operations like all-reduce.

Question 44Compute and GPU Fabric for AI Workloads

A cluster uses RoCEv2 for RDMA between GPU nodes. Which configuration is required to prevent packet reordering from degrading RDMA throughput?

  1. Increase MTU to 1500 bytes on every interface
  2. Set all RDMA traffic to the default best-effort class
  3. Disable ECN marking on all uplinks
  4. Enable Priority Flow Control with a lossless class for RDMA traffic Correct answer.

Explanation

RoCEv2 relies on a lossless Ethernet class, typically achieved with PFC on a dedicated priority, to avoid drops and reordering that harm RDMA performance. ECN and DCQCN then complement PFC for congestion control.

Question 45Compute and GPU Fabric for AI Workloads

Which statement describes the role of GPUDirect RDMA in an AI fabric?

  1. It lets a NIC transfer data directly to or from GPU memory without staging through host memory Correct answer.
  2. It virtualizes GPUs across multiple tenants
  3. It encrypts GPU memory contents in transit
  4. It replaces NVLink with Ethernet inside a node

Explanation

GPUDirect RDMA enables a network adapter to read and write GPU memory directly, bypassing host memory copies and reducing latency for inter-node transfers. This is central to efficient distributed training and inference data movement.

Question 46Compute and GPU Fabric for AI Workloads

An engineer observes that all-reduce time scales poorly as nodes are added, even though each link is lightly utilized. Which cause is most consistent with this symptom?

  1. The oversubscription ratio at the spine is high, causing contention on shared uplinks Correct answer.
  2. The dataset is stored on local NVMe instead of shared storage
  3. The collective is bound by per-link serialization rather than aggregate fabric capacity
  4. The GPUs are running at reduced clocks due to thermal throttling

Explanation

Lightly utilized individual links combined with poor scaling points to contention on oversubscribed shared uplinks where multiple leaves converge. Aggregate fabric capacity, not single-link speed, limits the collective in that case.

Question 47Compute and GPU Fabric for AI Workloads

Which congestion control mechanism is commonly paired with RoCEv2 to signal congestion using ECN marks?

  1. STP
  2. OSPF
  3. LACP
  4. DCQCN Correct answer.

Explanation

DCQCN is a congestion control algorithm designed for RoCEv2 that reacts to ECN marks to adjust sender rates. STP, LACP, and OSPF are Layer 2 or routing protocols unrelated to RDMA congestion control.

Question 48Compute and GPU Fabric for AI Workloads

What does an oversubscription ratio of 3:1 at a spine layer mean for an AI fabric?

  1. The total downlink capacity is three times the total uplink capacity toward the spine Correct answer.
  2. Each host gets three times its access bandwidth to the spine
  3. Each leaf has three redundant power supplies
  4. Three spines are reserved for failover only

Explanation

An oversubscription ratio compares downlink capacity to uplink capacity; 3:1 means the aggregate downlink bandwidth is three times the capacity available toward the spine. Higher ratios reduce cost but increase contention for collective traffic.

Question 49Compute and GPU Fabric for AI Workloads

Which Cisco technology allows a single physical GPU to be partitioned into multiple isolated instances for sharing among workloads?

  1. MIG Correct answer.
  2. LISP
  3. NAT
  4. VXLAN

Explanation

Multi-Instance GPU (MIG) partitions a supported GPU into isolated instances, each with dedicated memory and compute slices. VXLAN, LISP, and NAT are network functions with no GPU partitioning role.

Question 50Compute and GPU Fabric for AI Workloads

A team needs to verify that a GPU node can reach RDMA line rate to its peer. Which tool or method is most appropriate?

  1. Measure CPU utilization during a local disk benchmark
  2. Check the ARP table on both switches
  3. Run a ping flood between the hosts
  4. Use an RDMA bandwidth test such as perftest between the nodes Correct answer.

Explanation

RDMA bandwidth tests measure actual queue-pair throughput and latency between nodes, confirming the RDMA path works at expected rates. Ping and ARP checks only confirm basic IP reachability, not RDMA performance.