A 16-bit number format with the wide number range of FP32.
BF16 is like a kitchen scale with fewer decimal places. It still handles a giant turkey without freaking out.
It uses less memory and runs faster. AI training and answers often use it.
Quantization
BF16 uses lower precision but keeps a wide number range.
GPU
Modern GPUs support BF16 and speed up training and inference.
TPU
TPUs helped make BF16 common in large AI training.