Uniform Quantization

The defining feature of uniform quantization is a constant step size (Δ) between adjacent quantized values. This step size is calculated as Δ = (max - min) / (2^b - 1), where max and min define the clipping range and b is the bit-width. This uniform spacing simplifies the quantization and dequantization formulas to affine transformations: Q = round(r / Δ) + z and r' = (Q - z) * Δ, where Q is the quantized integer, z is the zero-point, and r' is the dequantized value.

• Example: For an 8-bit (b=8) representation of values in range [-1.0, 1.0], Δ = (1.0 - (-1.0)) / (255) ≈ 0.00784.
• This regularity is key to its hardware-friendly implementation.

Uniform quantization maps directly to efficient integer arithmetic on common hardware. The uniform step allows convolution and matrix multiplication operations to be performed entirely with integers, leveraging fast INT8 or INT4 units in CPUs, GPUs, and NPUs. The dequantization of the final result is a single, inexpensive floating-point operation.

• Performance Gain: Integer operations typically offer 2-4x speedup and significant power savings compared to equivalent FP32 operations due to reduced memory bandwidth and simpler logic circuits.
• This characteristic is the primary driver for its use in on-device inference and edge AI deployments.

Uniform quantization is implemented in two primary modes, defined by the zero-point (z) parameter.

• Symmetric Quantization: The quantized range is symmetric around zero (e.g., [-127, 127] for 8-bit). This forces z = 0, simplifying the arithmetic by eliminating zero-point subtraction during operations. It is optimal for data distributions (like weights after normalization) that are roughly symmetric.
• Asymmetric Quantization: The quantized range maps exactly to the tensor's observed min/max values (e.g., [0, 255]). The zero-point is non-zero, allowing a tighter fit for asymmetric distributions (like ReLU activations, which are all non-negative). This often reduces quantization error but adds an extra integer addition per operation.

The error introduced by uniform quantization is deterministic and bounded. The quantization error for any value is the difference between the original floating-point number and its dequantized representation. With uniform quantization, this error is uniformly distributed within ±Δ/2 for round-to-nearest.

• Granularity vs. Range Trade-off: For a fixed bit-width, a wider clipping range increases Δ, leading to coarser quantization and higher error for most values. A narrower range reduces Δ but increases clipping error for outliers that are saturated to the min/max limits.
• This predictable error profile makes analysis and calibration more straightforward compared to non-uniform schemes.

Determining the optimal clipping range [min, max] is a critical calibration step. This range defines Δ and directly impacts accuracy. Common methods include:

• Min-Max: Use the absolute observed min/max values from a calibration dataset. Simple but sensitive to outliers.
• Entropy Minimization (KL Divergence): Selects a range that minimizes the information loss between the original and quantized distributions.
• Mean-Squared Error (MSE) Minimization: Selects a range that minimizes the expected quantization error.
• Calibration is typically performed per-tensor (one range for an entire tensor) or per-channel (unique ranges for each output channel of a weight tensor), with per-channel offering higher accuracy.

The key distinction lies in the spacing of quantization levels.

• Uniform: Levels are evenly spaced (constant Δ). Advantages are computational simplicity and hardware-native support.
• Non-Uniform: Levels are unevenly spaced, allowing more levels in high-density regions of the value distribution (e.g., using logarithmic spacing). This can achieve lower quantization error for a given bit-width but requires complex, hardware-inefficient lookup tables or specialized operations for dequantization.

In practice, uniform quantization is overwhelmingly preferred for activation and weight quantization in production due to its efficiency, despite non-uniform schemes potentially offering better theoretical compression.

A uniform quantization scheme where the quantized range is symmetric around zero. This simplifies computation by setting the zero-point parameter to zero, meaning the integer value 0 directly represents the floating-point value 0. It's defined by a single scale factor applied to both positive and negative ranges. While computationally efficient, it can waste representational capacity if the original tensor's value distribution is not symmetric.

A uniform quantization scheme where the quantized range is mapped to the actual minimum and maximum values of the tensor. This uses both a scale and a non-zero zero-point parameter, providing a tighter fit to asymmetric data distributions (e.g., ReLU activations that are all non-negative). It typically yields lower quantization error for such data but adds a small computational overhead for zero-point correction during operations.

The two fundamental parameters that define the affine transformation for uniform quantization. They map between floating-point and integer domains.

• Scale (S): The ratio between the quantized integer range and the original floating-point range. Calculated as (float_max - float_min) / (quant_max - quant_min).
• Zero-Point (Z): The integer value that corresponds to the floating-point zero. It ensures that zero is quantized without error, which is critical for operations like padding. The transformation is: float_value = scale * (int_value - zero_point).

The numerical distortion or loss of information introduced when mapping a continuous range of floating-point values to a finite set of discrete integer levels. In uniform quantization, the primary source of error is rounding error from mapping to the nearest grid point. The maximum rounding error for a given step size (Δ) is ±Δ/2. This error propagates through the network and can accumulate, potentially degrading model accuracy. Techniques like calibration and quantization-aware training aim to minimize this error.

The process of analyzing a representative dataset (the calibration set) to estimate the optimal dynamic range for quantization. For static quantization, this step determines the fixed scale and zero-point for activations by observing their min/max values or other statistics (e.g., percentile) during a forward pass. Proper calibration is critical for minimizing clipping error (values outside the range) and rounding error, directly impacting final model accuracy post-quantization.

A contrasting scheme where quantization levels are not evenly spaced. This allows a higher density of levels in regions where the value distribution is more concentrated (e.g., near zero for weight distributions), potentially reducing quantization error for the same bit-width. However, it requires more complex, often non-linear, transformation functions and is generally less supported by standard integer arithmetic hardware compared to uniform quantization, which benefits from constant step size and simpler scaling.