How many bits are used for each number, and how they are arranged. Fewer bits means less memory and more speed, with some loss of accuracy. The names are cryptic: FP32 is full precision, BF16 and FP16 are half, FP8 and INT8 are quarter, and INT4 is what lets a large model run on a laptop.
Technical view
Floating point formats trade exponent bits, which give range, against mantissa bits, which give precision. BF16 keeps FP32’s exponent range with fewer mantissa bits, avoiding the overflow problems that make FP16 training require loss scaling. Integer formats need scale factors, applied per tensor, per channel or per group, with finer granularity preserving more quality. Weights tolerate lower precision than activations, and outlier channels dominate quantisation error.
Related
Part of AI Terminology Index, section: Making Models Smaller, Faster and Cheaper.