An exponent sets the binary scale of a floating-point value: value = significand × 2^exponent. It moves the representational grid, letting the same significand precision cover different magnitudes; unlike fixed-point, the spacing between values changes with scale. section 2.1
This adaptive range has a computational cost. Floating-point addition requires exponent alignment, and operations require normalization and rounding; floating-point also increases memory traffic and energy costs in inference. section 1.3.1
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
|
Toy floating-point examples show how a shared mantissa grid represents small and large values by changing the exponent. section 2.1





