Conceptual

Floating-Point Number Representation

IEEE 754 sign/exponent/mantissa layout, precision-range tradeoffs, rounding error, and the reduced-precision formats (fp16/bf16) used in ML.