Which process requires specialized hardware, large infrastructure, and a constant supply of energy?

• A. Inference
• B. Training a neural network
• C. Data labeling
• D. Hyperparameter tuning

Answer: (B)

Training a neural network requires substantial compute power because the learning process involves adjusting a vast number of parameters through backpropagation across many passes on large datasets. This work benefits greatly from specialized hardware like GPUs or TPUs that can handle parallel operations on millions or billions of numbers. To complete training in a practical time frame, you also need large-scale infrastructure—multiple machines, fast networking, substantial storage, and orchestration to coordinate distributed computation. All of this runs on a steady, reliable energy supply to keep the hardware powered and cooled, often for extended periods.

In contrast, producing inferences after training and labeling data don’t typically demand the same scale of continuous, high-intensity compute and power. Hyperparameter tuning can require additional compute, but the defining factor here is that the intense, sustained hardware and energy needs are most characteristic of the training phase.