liveBook · Manning

Step value

As explained in Quantization and Fast Inference

A step size is the spacing between adjacent representable levels in a uniform quantization grid.

Step size (S) is the spacing between adjacent levels on a uniform quantization grid. It sets resolution: smaller steps represent nearby values more precisely, while larger steps increase granular error. In fixed-point formats, allocating more bits to fractional values reduces the step size but narrows the representable range. section 2.1 section 2.3

Resolution and error

An 8-bit Q3.4 format has a step size of 1/16, or 0.0625. With a uniform quantizer using S = 0.05, values inside the grid have rounding error bounded by S/2; widening the grid reduces clipping but coarsens its steps. section 2.1 section 2.3

|error| ≤ S/2

Use the bound to estimate the largest granular error for values that remain inside the quantization grid. section 2.3

How Quantization and Fast Inference covers it

How deep it goes
Quantifies resolution with a Q3.4 example and an S/2 error bound, then relates step size to range and clipping tradeoffs.
Examples
Uses a Q3.4 step of 1/16 and a uniform quantizer with S = 0.05 to illustrate resolution and error.
What you can do after
Configure quantization resolution by balancing rounding error against clipping risk for the observed value range.
How it connects
Connects step size to fixed-point representation, Q-format, granular error, overload error, and calibration.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Quantization and Fast Inference. The section links point to the parts of the book it draws on.