Overview
Domain 1 is the architectural foundation of NCP-AIN and has three official objectives. The first asks you to describe an AI factory networking architecture and its components, such as GPUs, NVIDIA BlueField, scalable units and switches, and how the compute, storage, in-band and out-of-band networks are separated. The second covers rail-optimized topologies, where the same-numbered GPU NIC in every server connects to the same leaf switch. The third covers GPU-to-GPU communication: NVLink and NVSwitch inside a node or rack, GPUDirect RDMA across the fabric, and NCCL collectives that drive the traffic.
At 5%, this domain carries only a handful of questions, but its vocabulary appears throughout the Spectrum-X, InfiniBand and troubleshooting domains. NVIDIA has announced that lab exercises join this exam on November 1, so the delivery format may differ from older preparation material.
How it's tested
Questions here are descriptive rather than command-driven. They present a cluster layout or a traffic description and ask which component, network plane, topology or communication path applies. Expect wording about scalable units, rails, NVLink domains and east-west versus north-south traffic, with distractors that mix scale-up and scale-out technologies.
-
Identify the network plane
A traffic type such as checkpoint writes or BMC access is described, and you pick the compute, storage, in-band or out-of-band network.
-
Explain the rail
A cabling plan is shown and you decide whether it is rail-optimized and which GPU pairs reach each other in one switch hop.
-
Choose the GPU path
Two GPUs need to exchange data, and you select NVLink, PCIe peer-to-peer or GPUDirect RDMA over the fabric.
A question states that GPU 3 in server A must send a gradient shard to GPU 3 in server B on a rail-optimized cluster. Both NICs for GPU 3 land on the same rail leaf, so the traffic crosses one switch hop without reaching the spine. If GPU 3 had to reach GPU 5 in another server, NCCL PXN would first move data over NVLink to the right rail.
Objectives 3
Official objectives from the NVIDIA outline, in outline order.
1.1 Describe an AI factory networking architecture and its components (e.g., GPUs, NVIDIA BlueField, scalable units, switches)Building blocks and network planes of an NVIDIA AI factory.
- What it means
- An AI factory is a data center built to produce AI models and inference at scale. Its network combines GPU servers, BlueField DPUs or SuperNICs, leaf and spine switches and storage, arranged in repeatable scalable units such as the 32-node SU used by DGX SuperPOD reference designs.
- What to understand
- Know the separate planes: the east-west compute fabric (InfiniBand or Spectrum-X Ethernet) carries GPU traffic, a storage fabric serves datasets and checkpoints, an in-band management network handles provisioning, and an out-of-band network reaches BMCs and switch management ports. BlueField-3 DPUs typically handle north-south and infrastructure offload, while SuperNICs serve east-west GPU traffic.
- How it shows up
- Given a workload or device, identify which network plane or component handles it in the reference architecture.
- Common confusion
- Treating the DPU and the SuperNIC as interchangeable roles instead of north-south versus east-west functions.
1.2 Describe rail-optimized topologies for high-performance AI workloadsConnecting same-index GPU NICs across servers to shared leaf switches.
- What it means
- In a rail-optimized topology, NIC n of every server in a scalable unit connects to the same leaf switch, called rail n. With eight GPUs per server there are eight rails, so traffic between same-index GPUs needs only one switch hop.
- What to understand
- Collectives such as all-reduce mostly pair GPUs with the same local rank across nodes, so rails keep that traffic on a single leaf and reduce spine load and latency. Cross-rail traffic is handled by NCCL PXN, which moves data over NVLink inside the node to the NIC on the correct rail. Rail validation tools confirm cabling matches the plan.
- How it shows up
- A cabling diagram or traffic flow is described, and you decide how many hops it takes or whether the design is rail-optimized.
- Common confusion
- Assuming rail-optimized means one rail per server; a rail is per GPU index across servers.
1.3 Describe GPU-to-GPU communicationsNVLink, NVSwitch, GPUDirect RDMA and NCCL collectives.
- What it means
- This objective covers how GPUs move data to each other: NVLink and NVSwitch for high-bandwidth scale-up inside a server or an NVL72 rack, and GPUDirect RDMA for scale-out over InfiniBand or RoCE where the NIC reads and writes GPU memory directly.
- What to understand
- NCCL selects the transport automatically, using NVLink when peers share an NVLink domain and the network otherwise. GPUDirect RDMA bypasses host memory and CPU copies, lowering latency. SHARP offloads reductions into InfiniBand switches. Know the main collectives, all-reduce, all-gather, reduce-scatter and all-to-all, and that all-to-all is typical of mixture-of-experts models.
- How it shows up
- A data path between two GPUs is described, and you select the interconnect or library feature that carries it.
- Common confusion
- Believing GPUDirect RDMA uses NVLink; it is a NIC-to-GPU-memory path across the network.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
AI factory
A data center purpose-built to train and serve AI models at scale.
- Remember
- Networking is treated as part of the compute system.
- Tested as
- Architecture description questions.
- Trap
- Treating it as a generic enterprise data center.
-
Scalable unit (SU)
A repeatable block of GPU nodes and switches used to grow a cluster.
- Remember
- DGX SuperPOD reference designs use 32 nodes per SU.
- Tested as
- Sizing and design questions.
- Trap
- Confusing an SU with a single rack.
-
Compute fabric
The east-west network that carries GPU-to-GPU traffic between servers.
- Remember
- It is InfiniBand or Spectrum-X Ethernet, separate from storage.
- Tested as
- Network plane questions.
- Trap
- Sending storage or management traffic over it by default.
-
Out-of-band management
A separate network for BMCs and switch management ports.
- Remember
- It must survive failures of the data networks.
- Tested as
- Plane identification.
- Trap
- Mixing it up with in-band provisioning traffic.
-
BlueField-3 DPU
A data processing unit with Arm cores that offloads networking, storage and security.
- Remember
- Typically placed on the north-south path.
- Tested as
- Component role questions.
- Trap
- Assuming it is only a faster NIC.
-
SuperNIC
A network adapter optimized for east-west AI traffic in Spectrum-X.
- Remember
- Usually one per GPU, for example BlueField-3 or ConnectX-8 SuperNIC.
- Tested as
- Spectrum-X component questions.
- Trap
- Treating it as a north-south security device.
-
Rail
The set of same-index GPU NICs across servers connected to one leaf.
- Remember
- Same-rank GPU traffic stays one hop.
- Tested as
- Topology questions.
- Trap
- Thinking a rail belongs to one server.
-
PXN
NCCL PCIe-times-NVLink feature that relays data to the NIC on the right rail.
- Remember
- It avoids cross-rail spine traffic.
- Tested as
- Cross-rail traffic questions.
- Trap
- Assuming the switch fabric fixes cross-rail paths alone.
-
NVLink and NVSwitch
High-bandwidth scale-up links between GPUs in a server or NVL72 rack.
- Remember
- Scale-up, not scale-out.
- Tested as
- GPU path questions.
- Trap
- Calling NVLink part of the Ethernet fabric.
-
GPUDirect RDMA
Lets the NIC read and write GPU memory directly over the network.
- Remember
- Removes host memory bounce buffers.
- Tested as
- Data path questions.
- Trap
- Confusing it with GPUDirect Storage.
-
NCCL collectives
Library operations such as all-reduce and all-to-all that drive AI traffic.
- Remember
- Traffic is synchronized, so the slowest flow sets job time.
- Tested as
- Workload behavior questions.
- Trap
- Treating AI traffic like many small independent flows.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Scale-up vs. scale-out interconnects
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| NVLink / NVSwitch | GPU links inside a server or NVLink rack domain | Highest bandwidth, limited reach | Placing it in the leaf-spine fabric |
| InfiniBand compute fabric | Lossless, credit-based scale-out network | Managed by a subnet manager | Assuming it needs PFC |
| Spectrum-X Ethernet | RoCE fabric with Spectrum switches and SuperNICs | Ethernet operations with AI tuning | Treating it as ordinary best-effort Ethernet |
AI factory network planes
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Compute (east-west) | GPU-to-GPU collective traffic | Most bandwidth-sensitive | Sharing it with storage |
| Storage | Datasets and checkpoints | Separate fabric or converged by design | Ignoring checkpoint bursts |
| In-band management | Provisioning, job scheduling, user access | Runs on host OS interfaces | Calling it out-of-band |
| Out-of-band | BMC and switch console access | Independent of data planes | Routing it over the compute fabric |
DPU vs. SuperNIC roles
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| BlueField-3 DPU | Infrastructure offload with Arm cores | North-south, security, storage | Using it as the per-GPU NIC role |
| SuperNIC | AI-optimized east-west adapter | Congestion control and ordering for Spectrum-X | Expecting tenant firewall offload |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
NVLink spans the cluster
- What candidates pick
- Pick NVLink to connect GPUs in different scalable units.
- Why it looks right
- NVLink is the fastest GPU link.
- Why it's wrong
- NVLink is a scale-up interconnect within a server or NVLink rack domain; cross-node traffic in different SUs uses the InfiniBand or Ethernet fabric.
Clue to watch for The GPUs are in separate servers or SUs.
-
Rail equals server
- What candidates pick
- Describe each server as one rail.
- Why it looks right
- Rails sound like server uplinks.
- Why it's wrong
- A rail groups the same NIC index across all servers on one leaf, so each eight-GPU server touches eight rails.
Clue to watch for The diagram shows NIC 0 of many servers on one switch.
-
Cross-rail needs the spine
- What candidates pick
- Assume every cross-rail transfer must climb to the spine.
- Why it looks right
- Different leaves usually mean a spine hop.
- Why it's wrong
- NCCL PXN can move the data over NVLink to a local GPU on the destination rail first, keeping it on one leaf.
Clue to watch for The question mentions NCCL or intra-node NVLink relaying.
-
DPU on every GPU
- What candidates pick
- Choose BlueField-3 DPU mode as the per-GPU east-west adapter.
- Why it looks right
- BlueField is NVIDIA's flagship adapter.
- Why it's wrong
- The east-west role in Spectrum-X is the SuperNIC; DPU mode is aimed at north-south infrastructure offload.
Clue to watch for Wording about east-west GPU traffic and congestion control.
-
GPUDirect means storage
- What candidates pick
- Pick GPUDirect Storage when GPUs exchange data over the network.
- Why it looks right
- Both share the GPUDirect name.
- Why it's wrong
- GPUDirect Storage moves data between storage and GPU memory; GPU-to-GPU transfers across nodes use GPUDirect RDMA.
Clue to watch for Both endpoints are GPUs.
-
Average latency matters most
- What candidates pick
- Optimize for average flow latency.
- Why it looks right
- Average figures are common metrics.
- Why it's wrong
- Synchronized collectives wait for the slowest participant, so tail latency and effective bandwidth decide job time.
Clue to watch for The scenario mentions all-reduce or step time.
Answer framework
Use these steps to answer Domain 1 architecture questions about planes, rails and GPU data paths.
- Classify the traffic described as GPU collective, storage, provisioning or device management, which tells you the network plane involved.
- Decide whether the communicating GPUs share an NVLink domain; if so, scale-up NVLink applies, otherwise the scale-out fabric carries the data.
- For scale-out paths, check the rail layout and whether source and destination GPUs share the same local index and therefore the same leaf.
- Assign the adapter role correctly, with SuperNICs on east-west GPU traffic and BlueField DPUs on north-south and infrastructure offload.
- Choose the answer that reduces hops and tail latency for synchronized collectives, since the slowest participant determines how long each training step takes.
Practice question
Want more like this? The NCP-AIN practice test has questions for every domain.
Checklist
0 of 10
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Spend proportionate time here: five percent of the exam is only three or four questions, but the concepts carry into every other domain. Start with the NVIDIA DGX SuperPOD reference architecture and draw the four network planes yourself, labelling which devices connect to each. Add the scalable unit concept and note how many nodes, leaves and spines a single SU uses in the reference design.
Next, sketch a rail-optimized layout for two or three servers with eight GPUs each, and trace a same-rank transfer and a cross-rank transfer. Read the NCCL documentation on PXN so you can explain why cross-rail traffic rarely touches the spine. Then review NVLink, NVSwitch and the GB200 NVL72 rack so the scale-up domain is clearly separate from the fabric in your mind.
Finish with GPUDirect RDMA, SHARP and the main collectives. Write one sentence for each describing the traffic pattern it creates. These notes make the Spectrum-X and InfiniBand tuning objectives much easier, because you will know what problem each feature is solving.
FAQ
How much of NCP-AIN is AI Data Center Design and Optimization?
AI Data Center Design and Optimization is 5% of the exam, roughly three to four questions, with three official objectives.
Is the NCP-AIN exam being retired or changed?
No retirement is announced. NVIDIA's certification page says lab exercises will be added on November 1, so check the current format before you schedule.
What is a rail-optimized topology?
A layout where NIC n of every server connects to the same leaf switch, so GPUs with the same local index reach each other in one hop.
What is a scalable unit?
A repeatable building block of GPU servers and fabric switches. NVIDIA DGX SuperPOD reference designs commonly use 32 nodes per scalable unit.
How do GPUs in different servers exchange data?
Through GPUDirect RDMA over InfiniBand or Spectrum-X Ethernet, orchestrated by NCCL. NVLink handles GPUs inside the same server or NVLink rack.
What is the difference between a DPU and a SuperNIC?
BlueField-3 in DPU mode offloads infrastructure services on the north-south path, while a SuperNIC is tuned for east-west GPU traffic in Spectrum-X.
Do I need to memorize SuperPOD part numbers?
No official objective asks for part numbers. Focus on roles, planes, rails and data paths.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 1 knowledge
You have covered AI factory components, rail-optimized topologies and GPU-to-GPU paths. Check whether you can place traffic on the right plane and path under time pressure.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the NCP-AIN AI tutor Get any concept explained your way