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

Last checked October 22, 2026 against the official Cisco exam page.

1. Question 1AI network fabric design

   A cluster of GPU servers runs distributed training, and the team needs lossless, low-latency communication between nodes. Which network technology is most commonly used for AI back-end fabrics?

   1. ADial-up modem links that carry traffic over asynchronous PPP framing on a telephone line
   2. BSerial console links from a management workstation to an RS-232 port
   3. CPlain best-effort Ethernet forwarding with tail drop and no PFC or congestion control
   4. DRDMA over Converged Ethernet (RoCE) with lossless configuration, or InfiniBand
   Show answer and explanationHide answer

   Correct answer: D

   AI back-end fabrics need RDMA with lossless behaviour, provided by RoCE with congestion control or by InfiniBand. Best-effort Ethernet drops packets under load, and dial-up or serial links lack the bandwidth.

2. Question 2Monitoring and troubleshooting

   A distributed training job slows down after a switch firmware update. Which first check is most useful?

   1. ACompare interface error counters, congestion counters, and link utilization before and after the update
   2. BChange the job's name and scheduler queue label, then rerun training with the same random seed and dataset version
   3. CReinstall the GPU drivers and firmware stack without first checking the network fabric or the switch logs
   4. DRestart the monitoring agent on the head node and reset the dashboard refresh interval
   Show answer and explanationHide answer

   Correct answer: A

   Comparing error and congestion counters before and after the firmware change identifies network regressions. Monitor restarts, renaming jobs, and reinstalling drivers without checking the network do not isolate the cause.

3. Question 3Compute and GPU infrastructure

   A server with eight GPUs needs high-speed communication between those GPUs inside the chassis. Which technology provides this?

   1. AWi-Fi 6E radios between GPU cards
   2. BNVLink or NVSwitch within the server
   3. CSerial port sharing between GPU cards
   4. DUSB 1.1 hubs daisy-chained across GPUs
   Show answer and explanationHide answer

   Correct answer: B

   NVLink and NVSwitch provide high-bandwidth GPU-to-GPU communication inside the server. Wi-Fi, serial sharing, and USB 1.1 cannot match that bandwidth.

4. Question 4Orchestration and automation

   An operator wants to deploy the same AI cluster network configuration across many switches consistently. Which approach is best?

   1. ACopy configurations from random online sources and forum posts without peer review or lab testing
   2. BUse automation with templates or infrastructure as code, validated before deployment
   3. CChange settings when something breaks in production and users report the fault to the helpdesk
   4. DConfigure each switch by hand from memory during each maintenance window without a shared template
   Show answer and explanationHide answer

   Correct answer: B

   Templates or infrastructure as code with validation deliver consistent, reviewable configuration at scale. Manual memory-based configuration, unreviewed copies, and reactive changes produce drift.

5. Question 5Storage for AI workloads

   A training pipeline reads large datasets from shared storage, and GPUs wait for data. Which storage characteristic should be improved first?

   1. ASustained read throughput and parallel access to the dataset
   2. BThe storage array's label, asset tag, and the physical rack position
   3. CThe storage array's colour scheme and the paint finish on the front bezel
   4. DThe number of folders in the root directory of the share and their names
   Show answer and explanationHide answer

   Correct answer: A

   GPU starvation from slow data reads is best addressed by improving sustained parallel read throughput. Colour, labels, and folder counts do not affect data delivery.