Which technique is commonly used to speed up AI model training and inference on hardware...

Quantization reduces the numerical precision used to represent model weights and activations — for example, converting FP32 weights to INT8 — which decreases memory bandwidth requirements and allows hardware accelerators (GPU Tensor Cores, dedicated INT8 inference engines) to execute more operations per cycle, directly speeding up both training (in its mixed-precision form) and, especially, inference. Post-training quantization and quantization-aware training are the two dominant approaches, with the latter simulating quantization effects during training to better preserve accuracy at reduced bit-widths. NVIDIA's TensorRT relies heavily on quantization (alongside layer fusion and kernel auto-tuning) to accelerate deployed inference.

The distractors describe techniques that serve entirely different purposes: data augmentation (B) increases training data diversity to improve generalization, not computational speed — it typically adds preprocessing overhead rather than reducing it. Model enlargement (C) does the opposite of speeding up computation; larger models require more FLOPS and memory, increasing latency. Dropout (D) is a regularization technique applied during training to prevent overfitting by randomly zeroing activations — it has no role in inference-time speed (and is typically disabled at inference) and does not meaningfully accelerate training compute either.

Quantization is frequently paired with pruning and kernel/operator fusion as the three core techniques for hardware-accelerated performance optimization.