NCA-AIIO - AI Infrastructure (40% of the exam) - Section 2.10
Explain the purpose and benefits of a DPU in a data centre.
Explain how a Data Processing Unit (DPU) offloads networking, storage, and security tasks from the host CPU, freeing CPU and GPU cycles for AI computation. Recognise how NVIDIA BlueField DPUs enable software-defined networking, data compression, and infrastructure isolation without burdening the primary compute resources.
DPU
Practice question for this objective
lock_openFree sampleAI Infrastructuremedium
A virtualisation platform currently runs its software-defined networking layer, including the virtual switch that connects every guest to the fabric, on the host CPU. As guest density grows, the host spends an increasing fraction of its cycles forwarding packets between guests and the network. The team wants to relieve this without changing the network model the guests see. Which DPU capability most directly addresses the bottleneck?
APartitioning the host GPU into smaller instances so each guest receives a guaranteed slice of accelerator compute for packet forwarding.
BEnabling time-slicing on the network adapter so each guest's traffic is processed in a fixed scheduling quantum to even out CPU load.
CHardware-accelerating the virtual switch on the DPU so packet forwarding between guests and the fabric executes on the DPU rather than consuming host CPU cycles.check_circle Correct
DMigrating the guests to a separate physical host whose CPU has a higher core count to absorb the growing forwarding load.
A DPU offloads the software-defined virtual-switch data path into hardware, freeing host CPU cycles spent on packet forwarding. Software-defined networking running on the host means the CPU forwards every packet between guests and the fabric, and that cost rises with guest density. A DPU accelerates the virtual-switch data path in its own hardware, so forwarding no longer burns host cycles while the guests continue to see an unchanged virtual network.
Why A is wrong: Tempting because GPU partitioning is a real isolation technique, but packet forwarding is a networking task, not a GPU workload; carving up the GPU does nothing for virtual-switch overhead on the CPU.
Why B is wrong: Plausible because scheduling quanta manage contention, but time-slicing the adapter does not remove the forwarding work from the host CPU; the cycles are still spent, just reordered.
Why C is correct: Correct: the DPU can run the virtual-switch data path in hardware, so the per-packet forwarding work is offloaded from the host CPU while guests keep seeing the same software-defined network.
Why D is wrong: Tempting as a capacity fix, but adding host cores only relocates the same CPU-bound forwarding overhead; the DPU instead removes that overhead from the host entirely by offloading the data path.
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-AIIO bank for this domain.
The CPU performs collective gradient reductions, the GPU manages the storage data path, and the DPU runs the application's serial control logic.
errorWhy it is wrong: Tempting because each named task is real, but every assignment is misplaced: gradient reductions belong to GPUs, serial control logic to the CPU, and the storage data path is exactly the kind of infrastructure work a DPU offloads.
Multi-Instance GPU, which partitions one physical GPU into isolated instances so several data loaders can read in parallel.
errorWhy it is wrong: It is tempting because MIG improves sharing and parallelism, but partitioning a GPU into instances does nothing to remove the CPU bounce buffer on the storage-to-GPU path.
It increases the raw clock speed of the host CPU cores so that infrastructure services complete faster and release the cores sooner.
errorWhy it is wrong: Tempting because faster infrastructure processing sounds like the obvious goal, but a DPU does not alter the host CPU's clock; it removes the work entirely from those cores rather than speeding them up.