Page 1 of 183
Cisco Data Center AI Infrastructure (DCAI)
practice Exam – Questions with Answers &
Rationales| INSTANT PDF DOWNLOAD
Q1. Which type of AI workload involves training a model on a
large dataset to recognize patterns, requiring massive parallel
processing across hundreds or thousands of GPUs
simultaneously?
a) Model Inference
b) Model Training
c) Data Preparation
d) Model Evaluation
Answer: B
Rationale: Model training is the most computationally intensive
phase of the AI lifecycle. It involves processing vast datasets
through neural networks, requiring massive parallel computation
,Page 2 of 183
across GPU clusters. Training jobs can run for days or weeks and
generate tremendous east-west network traffic as gradients are
synchronized between GPUs .
Q2. A bank wants to deploy an AI system that instantly
approves or flags credit card transactions as they occur. Which
type of AI workload is most appropriate for this use case?
a) Model Training
b) Model Inference
c) Data Labeling
d) Feature Engineering
Answer: B
Rationale: Inference is the phase where a trained model makes
predictions on new, unseen data in real-time. Unlike training,
which is computationally heavy and time-consuming, inference
,Page 3 of 183
must be fast and responsive. For credit card fraud detection, sub-
second response times are critical .
Q3. Which workload characteristic distinguishes AI training
from traditional data center applications?
a) Predictable, steady-state traffic patterns
b) Bursty "incast" patterns where many nodes send data to a
single receiver simultaneously
c) Low bandwidth utilization
d) Minimal east-west traffic
Answer: B
Rationale: AI training creates unique "incast" traffic patterns
during collective communication operations (like all-reduce),
where many compute nodes send gradients to a single node or
synchronized group. This creates micro-bursts that can overwhelm
switch buffers without proper lossless configuration .
, Page 4 of 183
Q4. In a training workload using distributed data parallelism,
which protocol is commonly used for high-speed GPU-to-GPU
communication within the same server?
a) TCP/IP
b) UDP
c) NVLink
d) HTTP/2
Answer: C
Rationale: NVLink is NVIDIA's high-bandwidth, low-latency
interconnect that allows GPUs within the same server to share
memory and data directly, bypassing the PCIe bus. This is critical
for multi-GPU training efficiency .
Cisco Data Center AI Infrastructure (DCAI)
practice Exam – Questions with Answers &
Rationales| INSTANT PDF DOWNLOAD
Q1. Which type of AI workload involves training a model on a
large dataset to recognize patterns, requiring massive parallel
processing across hundreds or thousands of GPUs
simultaneously?
a) Model Inference
b) Model Training
c) Data Preparation
d) Model Evaluation
Answer: B
Rationale: Model training is the most computationally intensive
phase of the AI lifecycle. It involves processing vast datasets
through neural networks, requiring massive parallel computation
,Page 2 of 183
across GPU clusters. Training jobs can run for days or weeks and
generate tremendous east-west network traffic as gradients are
synchronized between GPUs .
Q2. A bank wants to deploy an AI system that instantly
approves or flags credit card transactions as they occur. Which
type of AI workload is most appropriate for this use case?
a) Model Training
b) Model Inference
c) Data Labeling
d) Feature Engineering
Answer: B
Rationale: Inference is the phase where a trained model makes
predictions on new, unseen data in real-time. Unlike training,
which is computationally heavy and time-consuming, inference
,Page 3 of 183
must be fast and responsive. For credit card fraud detection, sub-
second response times are critical .
Q3. Which workload characteristic distinguishes AI training
from traditional data center applications?
a) Predictable, steady-state traffic patterns
b) Bursty "incast" patterns where many nodes send data to a
single receiver simultaneously
c) Low bandwidth utilization
d) Minimal east-west traffic
Answer: B
Rationale: AI training creates unique "incast" traffic patterns
during collective communication operations (like all-reduce),
where many compute nodes send gradients to a single node or
synchronized group. This creates micro-bursts that can overwhelm
switch buffers without proper lossless configuration .
, Page 4 of 183
Q4. In a training workload using distributed data parallelism,
which protocol is commonly used for high-speed GPU-to-GPU
communication within the same server?
a) TCP/IP
b) UDP
c) NVLink
d) HTTP/2
Answer: C
Rationale: NVLink is NVIDIA's high-bandwidth, low-latency
interconnect that allows GPUs within the same server to share
memory and data directly, bypassing the PCIe bus. This is critical
for multi-GPU training efficiency .
