What does mixed-precision training refer to?

Mixed-precision training performs the bulk of computation — matrix multiplications and convolutions — in a lower-precision floating-point format (typically FP16 or BF16 on NVIDIA Tensor Cores) while maintaining a master copy of weights and accumulating certain sensitive operations (like loss scaling and gradient accumulation) in FP32 to preserve numerical stability. The result is substantially faster training throughput and reduced memory footprint, since lower-precision arithmetic runs at higher effective FLOPS on hardware with dedicated Tensor Cores, without a meaningful loss of final model accuracy when combined with techniques like dynamic loss scaling to prevent gradient underflow.

Note that option A's specific mention of "double-precision" (FP64) is not how mixed precision is practiced in modern deep learning — production mixed-precision training combines FP16/BF16 with FP32, not FP64, since FP64 offers no throughput advantage on Tensor Core hardware and is rarely used in training pipelines. Despite that imprecision in the option's wording, A is still the only choice capturing the correct underlying concept: combining multiple numeric precision levels within one training run. Options B, C, and D all misdescribe mixed precision as a *data-type* or *modality* strategy, confusing numerical precision (a performance/optimization concept) with data modality (a multimodal-data concept) — a distinction the exam tests directly.