A public-health inference network expands with new scalable units. Which set best names the networking building blocks that glue the AI factory together?
Select an answer to reveal the explanation.
Short Explanation
Picture a civic AI factory like a multi-story municipal garage: GPUs are the cars doing work, BlueField is the toll gear, scalable units are the floors you add, and switches are the ramps connecting everything. Leave any of those out and the expansion plan falls apart.
Full Explanation
A public-health inference network that expands with scalable units is glued together by a consistent set of AI factory networking building blocks: GPUs that execute models, NVIDIA BlueField DPUs that offload networking and security at the host, scalable units that package repeatable compute/network modules, and fabric switches that provide the leaf-spine (or rail) east-west paths. Naming that full set keeps clinic and epidemiology workloads portable as the municipal pod grows: each new scalable unit brings the same GPU, DPU, and switch attachment pattern rather than ad-hoc gadgets.
CPUs, ConnectX adapters, frontend networks, and management switches cannot deliver the kernel-bypass operations and sub-microsecond latencies that GPU collectives require as the factory glue. They may appear in general IT or break-glass paths, but they cannot deliver the bandwidth, deterministic east-west behavior, or scalable-unit attachment that public-health collectives need. NVLink domains, Quantum switches, InfiniBand gateways, and compute pods also fail the building-block test: they name specific interconnect domains, gateways, and pods rather than the required GPUs, BlueField DPUs, scalable units, and fabric switches, so they lack the host-side DPU offload and repeatable unit boundaries that bind the AI factory. Spectrum switches, storage arrays, retrieval services, and inference servers swap the AI factory glue for storage, retrieval, and campus-adjacent gear, and such traffic must remain logically separated from GPU fabric to avoid contention for bandwidth during collective operations.
Exam caveat: distractors often swap familiar municipal peripherals for real AI networking components—reject anything that omits GPUs, DPUs, scalable units, or fabric switches from the glue list. Operational check: when the public-health network adds a scalable unit, verify inventory and wiring call out GPU nodes, BlueField DPUs, unit boundaries, and fabric switch attachment in one design package, and reject expansion plans that rely on CPUs, ConnectX adapters, frontend networks, and management switches, on NVLink domains, Quantum switches, InfiniBand gateways, and compute pods, or on Spectrum switches, storage arrays, retrieval services, and inference servers as the interconnect.