Overview
Physical Layer Management has only two objectives, but both involve hands-on configuration of hardware that sits between the server and its workloads. The first is configuring and managing a BlueField Network Platform. That includes understanding the difference between running a BlueField-3 as a DPU with its own Arm cores and operating system and running it as a NIC or SuperNIC, accessing the device through rshim or its management interface, installing a BFB image with DOCA, and setting device parameters with tools such as mlxconfig. The second is configuring Multi-Instance GPU (MIG) on supported data center GPUs. That includes enabling MIG mode, creating GPU instances and compute instances from the available profiles, and exposing them to workloads.
At 5%, this domain contributes only about four questions, but the commands and concepts reappear in Domains 3 and 4.
How it's tested
Questions in this domain are short configuration scenarios. They describe a BlueField card that needs a mode change, a fresh image or remote access, or a GPU that needs to be split for several users, and they ask for the right command, the right order of steps or the expected result.
-
Pick the BlueField mode
A requirement for offload, isolation or plain networking is given, and you choose DPU mode or NIC mode.
-
Order the MIG steps
You arrange enabling MIG mode, resetting the GPU, creating GPU instances and creating compute instances in the right sequence.
-
Choose the profile
Memory and isolation needs are described, and you select the MIG profile or count of instances that fits.
A team wants seven isolated inference services on one H100 80GB GPU. Time-slicing shares the GPU but gives no memory isolation. MIG gives hardware-isolated slices, and the smallest profile allows up to seven instances. The engineer enables MIG mode, resets the GPU, creates seven 1g.10gb GPU instances with their default compute instances, then confirms them with nvidia-smi before scheduling workloads.
Objectives 2
Official objectives from the NVIDIA outline, in outline order.
2.1 Configure and manage a BlueField Network Platform.BlueField DPU and SuperNIC modes, imaging, access and configuration.
- What it means
- This covers setting up NVIDIA BlueField devices such as BlueField-3: choosing the operating mode, accessing the card through rshim or its out-of-band management port, installing a BlueField bootstream (BFB) image with the DOCA software, and setting firmware parameters with tools such as mlxconfig.
- What to understand
- In DPU mode the Arm cores run their own Linux and own the network path, so offloads such as OVS run on the card and the host can be isolated from network control. In NIC mode the card behaves like a ConnectX adapter and the Arm side is not in the data path. Mode and firmware parameter changes take effect only after a reset or power cycle, and the BFB image and firmware should match the DOCA release in use.
- How it shows up
- A goal such as isolation, offload or plain connectivity is given, and you choose the mode, tool or step.
- Common confusion
- Assuming a BlueField in NIC mode can still run offloaded services on its Arm cores.
2.2 Configure MIG (AI and HPC).Partitioning supported GPUs into isolated instances.
- What it means
- This covers Multi-Instance GPU on supported data center GPUs such as A100, H100 and newer: enabling MIG mode, creating GPU instances from fixed profiles, creating compute instances inside them, and making the instances available to containers, Kubernetes or Slurm.
- What to understand
- MIG splits a GPU into up to seven instances, each with dedicated memory, cache and compute slices, giving fault and performance isolation. Enabling MIG mode requires a GPU reset. A GPU instance owns memory, and compute instances inside it share that memory. Profiles such as 1g.10gb or 3g.40gb on an 80 GB GPU have fixed sizes and placement rules.
- How it shows up
- A sharing requirement is described, and you decide whether MIG fits and which steps or profile to use.
- Common confusion
- Confusing MIG isolation with time-slicing, which shares a GPU without isolating memory.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
DPU mode
BlueField Arm cores run their own OS and control the network path.
- Remember
- Enables offload and host isolation.
- Tested as
- Mode selection questions.
- Trap
- Expecting offload in NIC mode.
-
NIC mode
BlueField acts as a ConnectX-style adapter with the Arm side out of the data path.
- Remember
- Simpler, no on-card services.
- Tested as
- Mode selection questions.
- Trap
- Thinking it still runs OVS on Arm.
-
rshim
A host-side driver giving console and file access to the BlueField Arm system.
- Remember
- Used to push BFB images and reach the console.
- Tested as
- Access questions.
- Trap
- Confusing it with the BMC.
-
BFB image
The BlueField bootstream that installs the OS and DOCA software on the card.
- Remember
- Match it to the DOCA release.
- Tested as
- Imaging questions.
- Trap
- Installing a mismatched image and firmware.
-
mlxconfig
Tool to read and set non-volatile firmware parameters on NVIDIA adapters and DPUs.
- Remember
- Changes need a reset to apply.
- Tested as
- Configuration questions.
- Trap
- Expecting changes to apply instantly.
-
MIG mode
A per-GPU setting that allows partitioning into instances.
- Remember
- Enabling it needs a GPU reset.
- Tested as
- MIG order questions.
- Trap
- Creating instances before the reset.
-
GPU instance vs. compute instance
A GPU instance owns memory slices; compute instances divide its compute.
- Remember
- Compute instances share their GPU instance's memory.
- Tested as
- MIG structure questions.
- Trap
- Treating them as the same layer.
-
MIG profiles
Fixed slice sizes such as 1g.10gb, 3g.40gb and 7g.80gb on an 80 GB GPU.
- Remember
- Up to seven instances per GPU.
- Tested as
- Sizing questions.
- Trap
- Expecting arbitrary memory sizes.
-
Time-slicing
Sharing a GPU by taking turns without memory isolation.
- Remember
- It is not MIG.
- Tested as
- Sharing choice questions.
- Trap
- Claiming it gives fault isolation.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
BlueField operating modes
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| DPU mode | Arm OS owns the network path | Offload and host isolation | Forgetting it needs a BFB image and DOCA |
| NIC mode | Card acts as a network adapter | No Arm services in the data path | Expecting offloads to run |
| SuperNIC role | BlueField-3 tuned for east-west GPU traffic | Focus on compute fabric bandwidth | Confusing it with a storage controller |
GPU sharing methods
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| MIG | Hardware partitions with dedicated memory | Strong isolation, fixed sizes | Expecting it on every GPU model |
| Time-slicing | Workloads take turns on the whole GPU | No memory isolation | Calling it isolated |
| Whole GPU | One workload owns the GPU | Full performance, no sharing | Wasting capacity on small jobs |
BlueField access paths
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| rshim | Host-side console and image push | Works from the host PCIe link | Thinking it needs network access |
| OOB management port | Network access to the Arm OS | Reachable without the host | Confusing it with the server BMC |
| Host tools | mlxconfig and flint from the host | Change firmware parameters | Forgetting the reset after changes |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
NIC mode offload
- What candidates pick
- Run OVS offload on a BlueField set to NIC mode.
- Why it looks right
- The card hardware is the same.
- Why it's wrong
- In NIC mode the Arm cores are not in the data path, so on-card services do not handle traffic.
Clue to watch for The scenario needs on-card processing or host isolation.
-
Instant mlxconfig
- What candidates pick
- Expect a new mode or parameter to work right after mlxconfig runs.
- Why it looks right
- The command reports success.
- Why it's wrong
- Non-volatile parameters apply only after a firmware reset or power cycle.
Clue to watch for The new setting is not visible after the change.
-
Skip the reset
- What candidates pick
- Create GPU instances immediately after turning on MIG mode.
- Why it looks right
- The enable command returned no error.
- Why it's wrong
- MIG mode takes effect only after the GPU is reset or the node reboots, so instance creation fails until then.
Clue to watch for MIG mode shows as pending.
-
Time-slicing isolates
- What candidates pick
- Pick time-slicing for tenants that need memory isolation.
- Why it looks right
- It also lets several jobs share a GPU.
- Why it's wrong
- Time-slicing gives no memory or fault isolation; MIG does.
Clue to watch for The requirement says each tenant needs isolated memory.
-
Custom MIG sizes
- What candidates pick
- Request a MIG instance with an arbitrary memory amount.
- Why it looks right
- Other virtualization allows flexible sizes.
- Why it's wrong
- MIG uses fixed profiles with set compute and memory slices and placement rules.
Clue to watch for The answer options list specific profile names.
-
rshim equals BMC
- What candidates pick
- Use the server BMC to install a BFB image on the BlueField.
- Why it looks right
- Both are management paths.
- Why it's wrong
- BFB images are pushed through rshim or the BlueField's own management path, not the host BMC.
Clue to watch for The task is imaging the DPU's Arm OS.
Answer framework
Use these steps for any BlueField or MIG configuration scenario in this small domain.
- Identify whether the scenario is about the BlueField platform or about partitioning a GPU, since the tools do not overlap.
- For BlueField, decide the required mode from the need for offload or isolation, then pick rshim, the management port or host tools for access.
- Remember that BlueField firmware parameter changes and image installs finish only after a reset, and versions should match the DOCA release.
- For MIG, confirm the GPU supports it, enable MIG mode, reset the GPU, then create GPU instances followed by compute instances.
- Pick the MIG profile from the memory and isolation need, and confirm the result with nvidia-smi before handing instances to a scheduler.
Practice question
Want more like this? The NCP-AII 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
This domain is small, so plan a short, practical review rather than deep study. Start with BlueField. Read the BlueField-3 and DOCA host installation documentation listed on the certification page and build a one-page summary of operating modes, access paths and imaging. Make sure you can say where software runs in each mode and why a reset is needed after a firmware parameter change.
Then spend most of the remaining time on MIG, ideally on a real or cloud GPU that supports it. Enable MIG mode, reset the GPU, list available profiles, create a mix of instances, destroy them and recreate a different layout. Write down the profiles for an 80 GB GPU and note which combinations fit.
Finally, connect both topics to the rest of the exam. BlueField firmware checks appear in Domain 4, DOCA driver installation appears in Domain 3, and MIG changes how Slurm and containers request GPUs. A handful of practice questions is enough to confirm this domain, so do not let it take time from the two heaviest domains.
FAQ
How much of the NCP-AII is Physical Layer Management?
Physical Layer Management is 5% of the NCP-AII, the smallest domain, which works out to about four questions on a 70 to 75 question exam.
What BlueField skills does the exam expect?
Configuring and managing a BlueField Network Platform: choosing DPU or NIC mode, reaching the card through rshim or its management port, installing BFB images with DOCA, and setting parameters with mlxconfig.
Which GPUs support MIG?
MIG is supported on data center GPUs starting with the Ampere A100 and including Hopper and newer models such as H100 and H200. It is not available on every GPU, so check support before planning partitions.
Why does MIG need a GPU reset?
Enabling MIG changes how the GPU is partitioned internally. The setting stays pending until the GPU is reset or the node reboots, and only then can instances be created.
What is the difference between a GPU instance and a compute instance?
A GPU instance owns a set of memory and compute slices. Compute instances are created inside a GPU instance and share its memory while splitting its compute.
Why is this called Physical Layer Management?
NVIDIA groups these objectives as configuring the hardware layer that workloads see: the BlueField network platform and the physical partitioning of GPUs.
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 2 knowledge
You have covered BlueField operating modes, imaging and configuration, plus MIG setup and profiles. Check whether you can order the steps correctly in a short scenario.
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-AII AI tutor Get any concept explained your way