A mantissa stores a floating-point number’s significant digits. In the representation value = mantissa × 2^exponent, the exponent sets the scale; the mantissa carries the detail at that scale. section 2.1
This moving scale lets one format represent both small and large values: a fixed-point grid fine enough for small numbers may overflow on large ones, while a grid sized for large values can erase small ones. Mantissa precision is therefore distinct from the range enabled by the exponent. section 2.1
Quantizing a toy mantissa
With four fractional bits, the mantissa is rounded to multiples of 1/16. For 0.10 ≈ 1.60 × 2⁻⁴, rounding 1.60 to 1.625 gives a reconstructed value of 0.1015625. section 2.1
Table 2.2 Quantize only the mantissa to the nearest multiple of 1/16
|
True Value
|
(Approx) mantissa × 2^exp
|
Quantized mantissa
|
Reconstructed value
|
|
0.01
|
1.28 × 2⁻⁷
|
1.3125 × 2⁻⁷
|
0.01025390625
|
|
0.10
|
1.60 × 2⁻⁴
|
1.6250 × 2⁻⁴
|
0.1015625
|
|
1.00
|
1.00 × 2⁰
|
1.0000 × 2⁰
|
1.0
|
|
3.00
|
1.50 × 2¹
|
1.5000 × 2¹
|
3.0
|
Compare the original and reconstructed values to see how mantissa rounding affects numbers at different exponents. section 2.1
Floating-point’s adaptive representation supports wide magnitude ranges, but inference may not use all its precision. Its variable spacing, exponent alignment, normalization, and rounding also contribute to computation and data-movement costs. section 1.3.1






