// CISCO DATA CENTER AI INFRASTRUCTURE · 300-640 DCAI

Cisco DCAI practice test — build and run the AI fabric

DCAI is Cisco's new AI infrastructure specialist exam and a CCNP Data Center concentration, and prep for it barely exists yet. CrushCert gives you 100 adaptive practice questions written to the official 300-640 topics — lossless Ethernet for RoCEv2 (PFC, ECN, ETS, DCQCN), GPU compute on UCS, storage for training and checkpoints, power and cooling, Nexus Dashboard, APIC, Intersight and Hyperfabric — with a written explanation on every answer, hands-on labs, and a full timed mock exam.

Start studying free See what's on the exam

7 days free · No card required · Built to pass you on the first try

100
practice questions
90
minutes on exam
4
exam domains
$15/mo
Single exam

How CrushCert gets you exam-ready

WHAT SETS US APART

Written to the 300-640 exam topics

Questions are weighted exactly like the real exam — 20/30/30/20 across the four domains — so the mix you practice is the mix you'll face, with design and deployment getting the 60% they deserve.

Adaptive practice

The algorithm tracks every question you miss and serves more of those topics, so your weak areas get more reps and every session moves your score.

Realistic mock exams

Full-length, timed tests that mirror the real exam — 60 questions, 90-minute countdown, question navigator, mark-for-review, and an 80% pass line so you know you're ready before exam day.

What's on the Cisco DCAI exam

Implementing Cisco Data Center AI Infrastructure (300-640 DCAI) is a 90-minute exam that costs $300 USD. It earns the Cisco Certified Specialist – Data Center AI Infrastructure credential and counts as a concentration exam for CCNP Data Center. Cisco does not publish the question count or passing score. The four domains and weightings:

DomainWeight
AI fundamentals and applications20%
AI infrastructure components and architecture30%
AI infrastructure deployment and data management30%
AI infrastructure operations and troubleshooting20%

CrushCert's question bank maps to every domain. Topic coverage includes:

PFC, ECN & ETSRoCEv2 & DCQCNNX-OS QoS for AIGPU, NVLink & GPUDirectUCS & Intersight policiesNexus Dashboard & APICHyperfabric AI & AI PODsStorage for training & checkpointsPower, cooling & PUEHybrid AI & data gravityNCCL & RDMA benchmarksTelemetry & troubleshooting

Train on every domain above

Adaptive practice questions and hands-on labs mapped to these exact domains. 7 days free, no card required.

Start free trial

Sample Cisco DCAI practice questions

Sample — AI Infrastructure Deployment
In a DCQCN-based RoCEv2 deployment, a Nexus spine sets the ECN CE bit on packets in a congested queue. What happens next so that the sender actually reduces its rate?
A) The spine sends a pause frame toward the sender's leaf, which propagates PFC hop by hop to the sending NIC
B) The receiving NIC sends Congestion Notification Packets (CNPs) to the sending NIC, which lowers its injection rate ✓
C) The spine drops the marked packets so that the RDMA transport retransmits at a lower rate after a timeout
D) The receiving leaf switch generates a CNP on behalf of the server and forwards it to the sending leaf

ECN marking is only a signal. The NIC that receives marked packets answers with CNPs, and the sending NIC's DCQCN logic cuts its rate. PFC pauses are the last-resort backstop, not the ECN reaction, and marked packets are never dropped. Every CrushCert question includes an explanation like this.

Sample — AI Infrastructure Deployment
An AI fabric is a routed leaf-spine with Layer 3 links between every switch. RoCE is classified on CoS 3 at the leaf and handled correctly there, but at the spine and remote leaf it is treated as best effort. What should change?
A) Enable PFC on the spine ports so the CoS value is restored on egress
B) Add a 'set cos 3' action in the egress policy-map on each leaf
C) Classify and trust on DSCP instead, because CoS is a Layer 2 field that is lost at routed hops ✓
D) Disable VXLAN on the AI fabric because the encapsulation strips the DSCP value

CoS lives in the 802.1Q header and does not survive a routed hop. End-to-end classification in a Layer 3 fabric has to key on DSCP, carried in the IP header, and the NICs must mark it consistently.

Sample — Components & Architecture
A data scientist asks why a model that fits in a server's 2 TB of host RAM still fails to train on a GPU with 80 GB of memory. Which explanation is correct?
A) GPU memory mainly caches input batches, so the dataset itself must exceed 80 GB
B) Parameters, gradients, optimizer state, and activations must reside in GPU HBM during training; host RAM is not where the GPU computes ✓
C) Host RAM has higher bandwidth than HBM, so the driver allocates the model there and the GPU times out
D) The GPU can address host RAM over PCIe at near-HBM speed, so this points to a driver defect

The GPU computes on what is in its own high-bandwidth memory. Model, optimizer state, and activations all have to fit there, which is why large models are split across GPUs and nodes instead of relying on host RAM.

Sample — Operations & Troubleshooting
Before handing a new 32-GPU training cluster to the data science team, an engineer runs the NCCL all_reduce_perf test across all nodes. What does this test primarily measure?
A) The peak floating-point throughput of each GPU under a sustained matrix-multiply workload
B) Collective-communication bandwidth achieved between GPUs across the cluster ✓
C) The read and write IOPS of the shared checkpoint storage
D) The TCP throughput of the management network between nodes

NCCL tests exercise the collectives distributed training depends on, so the result is the bus bandwidth the GPUs achieve through NVLink and the back-end fabric. It is the single best commissioning baseline for an AI fabric.

Sample — AI Fundamentals
A healthcare provider must keep patient data in-country on infrastructure it controls, but researchers want elastic GPU capacity for experiments on synthetic, non-sensitive data. Which deployment model BEST fits?
A) Public cloud in the nearest in-country region for everything, with encryption at rest
B) Hybrid: train on patient data on-premises and burst synthetic-data experiments to a public cloud ✓
C) Edge AI nodes at each clinic so no patient data leaves the building
D) A single on-premises cluster sized for peak experimental demand

Hybrid matches the two different requirements: sensitive data stays on controlled on-prem infrastructure, and non-sensitive experiments use elastic cloud GPUs. Sizing on-prem for peak wastes capacity, and edge nodes do not provide elastic training capacity.

Why learners pick CrushCert

You practice the way the exam tests you. DCAI is scenario-heavy — "a training job slowed down after a new leaf was added, what is the MOST likely cause?" — not definition recall. Every CrushCert question is written in that style, every explanation tells you why the other three choices are wrong, and hands-on labs drill the judgment calls — naming the lossless-Ethernet mechanism, matching workloads to what they need, picking the right Cisco tool, and working a RoCEv2 troubleshooting ticket from counters to fix.

Pay for one exam or all of them. Cisco DCAI alone is $15/month; the Professional plan is $24/month and covers all 17 exams — CCNP Data Center, CCNP Enterprise, CCNP Security, AWS Solutions Architect, CySA+ and more — so if you're stacking certs, it's all included.

You always know your readiness. A live readiness score blends your quiz accuracy, mock results, and lab performance, and points you to exactly what to study next.

A guided plan tells you what to study today. Guided Learning builds a short daily plan from your performance — a warm-up review, a lesson on your weakest topic, and flashcards, all in about 20 minutes, with spaced repetition so material comes back right before you'd forget it.

Keep learning: Cisco data center path

CORE EXAM

CCNP Data Center (DCCOR) practice test

The 350-601 core exam DCAI pairs with for CCNP Data Center — Nexus, ACI, UCS, storage networking and automation.

 RELATED

CCNP Enterprise practice test

Coming from the enterprise side? ENCOR covers the routing, QoS and automation fundamentals DCAI builds on.

 STUDY PLAN

Cisco DCAI study plan

A day-by-day 3-week or 6-week schedule, plus cheat sheets for lossless Ethernet and the counters that diagnose a slow AI fabric.

 AI FOUNDATIONS

AWS AI Practitioner practice test

New to AI workloads? The foundational AWS AI exam explains training, inference, RAG and generative AI from the ground up.

Not sure which cert comes next? Take the two-minute Which certification should I take? quiz.

Ready to crush Cisco DCAI?

Adaptive questions, hands-on labs, explanations on every answer, and full mock exams — start free for 7 days, no card required.

Start studying free

Cisco DCAI FAQ

What is the Cisco DCAI 300-640 exam?

300-640 DCAI (Implementing Cisco Data Center AI Infrastructure) is a 90-minute Cisco exam that earns the Cisco Certified Specialist – Data Center AI Infrastructure credential. It also counts as a concentration exam toward CCNP Data Center, so you can pair it with the DCCOR 350-601 core exam. It costs $300 USD.

How many questions are on the DCAI exam and what is the passing score?

Cisco does not publish question counts or passing scores for its exams. CrushCert's DCAI mock exams use 60 questions in 90 minutes with an 80% pass line, the same settings we use for the other CCNP-level Cisco exams.

What does the DCAI exam cover?

Four domains: AI fundamentals and applications (20%), AI infrastructure components and architecture (30%), AI infrastructure deployment and data management (30%), and operations and troubleshooting (20%). In practice that means lossless Ethernet for RoCEv2 (PFC, ECN, ETS, DCQCN), GPU compute on Cisco UCS, storage for training and checkpoints, power and cooling, and Cisco tools such as Nexus Dashboard, APIC, Intersight, and Hyperfabric AI.

Should I take DCCOR or DCAI first?

If you are working toward CCNP Data Center, most people take the DCCOR 350-601 core exam first, because DCAI assumes you already know Nexus, NX-OS QoS, and UCS basics. If you only want the AI Infrastructure specialist badge, DCAI on its own is fine. CrushCert covers both on one plan.