One number.
Different realities.
Explore how AI number formats trade precision for space.
Type a value. See exactly where every bit goes.
Decimal input · nearest rounding, ties to even · INT values saturate on overflow
Compare formats
Most significant bit firstBuild a format
Floating-point layouts use one sign bit, a biased exponent, subnormals, and reserved infinity/NaN encodings. Up to 64 total bits.
Range vs. precision
Exponent bits control scale. Fraction bits control detail. BF16 keeps FP32’s exponent width with fewer fraction bits; FP16 spends more of its bits on precision.
Reading the memory
Each box is one stored bit. Expand a format to see its formula, rounding error, and byte order. The leading 1 in a normal floating-point value is implicit: it is never stored.
A note on INT8
This is a direct integer conversion. AI quantization often adds a scale and zero point; those are separate from the underlying INT8 bit pattern.