Cisco DCAI study plan: 3-week and 6-week schedules
Here's a concrete, day-by-day plan for Cisco DCAI (300-640, Implementing Cisco Data Center AI Infrastructure): a 3-week track if you've passed DCCOR or work on Nexus and UCS every day, and a 6-week track if the data center is newer to you. Follow either one and finish consistently scoring 85%+ on practice exams before you book.
DCAI is not an AI theory exam. It's an infrastructure exam about the networks, servers, storage and power that AI workloads run on. The questions sound like a ticket queue: why a training job slowed down after a change, which fabric feature keeps RoCEv2 traffic from dropping, how much power and cooling a GPU rack needs, and which Cisco tool you'd open to fix it. This plan is built around those calls.
The exam you're planning for
| What | Detail |
|---|---|
| Exam code | 300-640 DCAI |
| Time | 90 minutes |
| Questions / pass score | Not published by Cisco |
| Earns | Specialist – DC AI Infrastructure |
| Also counts toward | CCNP Data Center (with DCCOR) |
| Domain | Weight |
|---|---|
| AI fundamentals and applications | 20% |
| AI infrastructure components and architecture | 30% |
| AI infrastructure deployment and data management | 30% |
| AI infrastructure operations and troubleshooting | 20% |
Architecture and deployment together are 60% of the exam, so both plans give them the most days. Because Cisco doesn't publish the passing score, aim higher than you would on other exams: 85% on practice tests, not 75%.
Pick your track
| You | Track |
|---|---|
| Passed DCCOR, or work on Nexus / UCS day to day | 3 weeks · ~1.5–2 hrs/day |
| Network or server background, newer to the data center | 6 weeks · ~1 hr/day |
| Going for CCNP Data Center | Start with DCCOR first |
The daily routine (both tracks)
Every study day follows the same shape:
- 20–30 minutes — learn: one topic from that day's list (the official exam topics, Cisco design guides and white papers, a course video, or your notes).
- 30–60 minutes — practice questions on that topic, plus a few review questions from earlier days.
- 10 minutes — review misses: read the explanation for every wrong answer and say out loud why the right answer wins and what each wrong one would actually do.
The 3-week plan (experienced track)
Week 1 — AI fundamentals (20%) and architecture (30%)
- Day 1: AI workloads from the infrastructure side: training vs fine-tuning vs inference, and what each needs from compute, network and storage. Why training is east-west heavy and inference is latency-sensitive.
- Day 2: How distributed training talks: data and model parallelism, collective operations like all-reduce, and why one slow link slows the whole job. GPU interconnects: NVLink inside a server vs the back-end network between servers.
- Day 3: Fabric design: front-end vs back-end (GPU) networks, non-blocking leaf-spine, rail-optimized designs, InfiniBand vs Ethernet trade-offs, and RoCEv2 (RDMA over UDP/IP, routable).
- Day 4: Compute and storage: UCS GPU servers, NICs and DPUs, storage for datasets and checkpoints, NVMe and NVMe over Fabrics, parallel file systems.
- Day 5: Facilities: power density per rack, PUE (total facility power ÷ IT power, closer to 1.0 is better), and when air cooling stops being enough and liquid cooling comes in. Then Cisco's validated stacks: AI PODs and Hyperfabric AI.
- Weekend: A mixed quiz on domains 1 and 2. Anything under ~75%, re-read that topic.
Week 2 — Deployment and data management (30%)
- Day 1: Lossless Ethernet end to end: PFC (802.1Qbb), ETS (802.1Qaz), ECN marking with WRED thresholds, and DCQCN with congestion notification packets. Learn the cheat sheet below until it's automatic.
- Day 2: NX-OS QoS for RoCEv2: classifying RoCE traffic into a no-drop class, the network-qos policy (pause and MTU), queuing policies, and enabling PFC on interfaces. Know what show interface priority-flow-control and show queuing interface tell you.
- Day 3: UCS and Intersight: server and domain profiles and the policies inside them (BIOS, power, boot, LAN connectivity/vNICs, QoS system classes, NTP), and how Intersight manages them from the cloud.
- Day 4: Fabric management tools: what Nexus Dashboard does (fabric controller and insights), what APIC does for ACI fabrics, and where Hyperfabric fits. Expect "which tool would you use" questions.
- Day 5: Data management: data pipelines, moving and staging datasets, checkpoint traffic, and keeping storage traffic from fighting with GPU traffic.
- Weekend: Domain 3 quiz plus a re-drill of misses.
Week 3 — Operations and troubleshooting (20%), then mock exams
- Day 1: Benchmarks and what each one proves: MLPerf for whole-system performance, NCCL tests for all-reduce bandwidth between GPUs, RDMA perftest (for example ib_write_bw) for a single link, iperf for plain TCP, fio for storage.
- Day 2: Monitoring and telemetry: streaming telemetry, Nexus Dashboard insights, GPU utilization, and the "read the counters" cheat sheet below.
- Day 3: Full timed mock exam (90 minutes). Score it and list every miss by domain.
- Day 4: Drill your two weakest domains.
- Day 5: Second full mock. You want 85%+.
- Exam day −1: Light review only — the two cheat sheets and the tool list. Stop early. Sleep.
Cheat sheet 1: the lossless-Ethernet toolkit
A big share of DCAI questions turn on which mechanism does what. Most wrong answers are one of these features applied to the wrong problem:
| The scenario needs… | Pick |
|---|---|
| Stop drops for one traffic class without pausing the whole link | PFC (802.1Qbb) |
| Guarantee each traffic class a share of link bandwidth | ETS (802.1Qaz) |
| Warn senders early, before queues fill and pauses start | ECN marking (WRED thresholds) |
| RoCEv2 senders slow down when they see congestion | DCQCN (ECN + CNPs) |
| RDMA traffic that can be routed across a leaf-spine fabric | RoCEv2 (UDP port 4791) |
| Fast GPU-to-GPU traffic inside one server | NVLink |
Watch for the classic trap: "enable PFC on everything." PFC belongs on the no-drop class that carries RoCE traffic. Pausing every class spreads congestion through the fabric instead of fixing it, which is why ECN and DCQCN do the first round of the work.
Cheat sheet 2: read the counters
Troubleshooting questions usually hand you a symptom or a counter and ask what it means:
| You see… | Think |
|---|---|
| Drops in the no-drop class | PFC/QoS not set on that hop |
| Pause frames climbing on one port | Congestion downstream of it |
| Lots of ECN marks / CNPs, no drops | Congestion control working |
| CRC errors on one link | Bad optic, cable or port |
| Low NCCL all-reduce bandwidth | Find the slow node or link |
| Problems right after adding a switch | Compare its QoS and MTU |
The 6-week plan (newer-to-the-data-center track)
Same topics, more room. Do the same daily routine at about an hour a day.
- Week 1: Data center basics you'll lean on: leaf-spine fabrics, VLANs vs VXLAN overlays, QoS classes and queues, and what UCS servers and fabric interconnects are. Then AI workloads: training vs inference and what each needs.
- Week 2: Finish AI fundamentals (distributed training, collectives, NVLink vs the back-end network), then start architecture: front-end vs back-end networks, rail-optimized designs, InfiniBand vs Ethernet, RoCEv2.
- Week 3: Finish architecture: GPU servers, NICs and DPUs, storage for AI, power, PUE and cooling, AI PODs and Hyperfabric AI. End with a domain 1–2 quiz.
- Week 4: Lossless Ethernet and NX-OS QoS. Take two days on PFC, ETS, ECN and DCQCN alone — it's the most-tested idea on the exam.
- Week 5: UCS policies and Intersight, Nexus Dashboard, APIC and Hyperfabric, data pipelines. End with a domain 3 quiz.
- Week 6: Benchmarks, monitoring and the counters cheat sheet, then two full timed mocks with a weak-spot drill in between. Book the exam once you clear 85% twice.
The final 48 hours (both tracks)
- Two days out: last full mock. If it's 85%+, you're done learning — trust it.
- One day out: a 20-minute skim of both cheat sheets and the tool list (Nexus Dashboard, APIC, Intersight, Hyperfabric). No new topics, no full exams.
- Exam day: you have 90 minutes. If a long scenario eats more than three minutes, pick your best answer, mark it, and come back with the time you've saved.
Run this plan on CrushCert
Adaptive DCAI practice questions with an explanation on every answer for the daily reps, hands-on labs for the fabric and troubleshooting calls, full timed mock exams for the final week, and a readiness score that tells you when to book. 7-day free trial, no card required.
Start the DCAI plan →Before and after DCAI
If you're going for CCNP Data Center, DCCOR is the required core exam and DCAI counts as your concentration. If AI concepts are the gap, the foundational AWS AI Practitioner covers models, training and inference from the ground up. Not sure which direction fits? Take the two-minute which certification should I take quiz.
Frequently asked questions
How long does it take to study for Cisco DCAI 300-640?
About 3 weeks at 1.5 to 2 hours a day if you already passed DCCOR or work on Nexus and UCS. About 6 weeks at an hour a day if data center networking is newer to you. Either way, finish with two full timed mock exams at 85% or higher.
Should I take DCCOR before DCAI?
If you want the CCNP Data Center certification, yes: you need the DCCOR (350-601) core exam plus one concentration exam such as DCAI. DCAI also stands on its own and earns the Cisco Certified Specialist – Data Center AI Infrastructure credential, but it assumes you already know data center networking and compute.
Do I need GPU hardware to study for DCAI?
No. The exam tests design, deployment and troubleshooting decisions, not hands-on GPU tuning. Knowing what PFC, ECN, ETS and DCQCN do, which counters to read, and which Cisco tool manages what will carry you much further than lab hardware.
What practice-exam score means I'm ready for DCAI?
Cisco does not publish a passing score, so aim high: 85% or better on two full timed practice exams in a row. That margin covers the unknown cut score and test-day nerves.