AI0-001 AI Infrastructure and Technologies Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Use FP32 precision

  Why it's wrong here

  Using FP32 precision requires 32-bit floating-point arithmetic, which demands significantly higher memory bandwidth and computational throughput than lower-precision formats. On resource-constrained edge devices with minimal power budgets, this precision level directly increases energy consumption and may exceed available RAM or processor capabilities, making real-time inference infeasible. It is tempting because FP32 offers maximum numerical accuracy and is the default for training models, so developers often assume it preserves model fidelity. In a cloud-based deployment with ample power and cooling, FP32 would be the correct choice to avoid any accuracy degradation from quantisation.

• ✓

  Model quantization (INT8)

  Why this is correct

  INT8 quantization shrinks weights and activations to 8-bit integers, cutting memory footprint and enabling integer arithmetic that draws far less power than FP32. This directly satisfies the stem's constraint of running on resource-constrained edge devices with minimal power consumption.

• ✗

  Increase the number of layers

  Why it's wrong here

  Adding layers increases parameters, computation and memory, directly raising power consumption and latency on constrained devices. It is tempting because deeper networks can improve accuracy, but that suits powerful servers, not edge hardware where the goal is reducing model size and energy use.

• ✗

  Use a larger batch size

  Why it's wrong here

  Larger batches raise peak memory and buffer requirements, worsening latency and power draw on constrained edge devices. It is tempting because large batches improve throughput on datacentre GPUs, but that scenario targets server-side inference, not battery-powered sensors needing per-sample, low-power execution.