Hybrid consensus uses symmetric quantization for weights and asymmetric quantization for activations. It keeps weight arithmetic efficient while fitting the activation grid to the data, particularly for one-sided ReLU outputs. section 2.4 section 2.6
Weights typically cluster around zero, so a centered grid fits them with little wasted range. Activations can be one-sided or shifted; an asymmetric grid uses a calculated zero-point to cover their range more effectively. Its zero-point can still be zero, as with a nonnegative ReLU range. section 2.4 section 2.5
Table 2.4 Quantization scheme comparison
|
Feature
|
Symmetric
|
Asymmetric
|
Hybrid Consensus
|
|
Zero Point (Z)
|
Fixed at 0
|
Calculated from min/max
|
Zw=0, Zx
calculated
|
|
Grid Constraint
|
Centralized
([−α,α])
|
Flexible ([α,β]
|
Weights central, activations flexible
|
|
Compute Overhead
|
Lowest. Raw dot product.
|
Highest.
Requires ∑qx and ∑qw.
|
Low.
Removes ∑qx overhead.
|
|
ReLU Efficiency
|
Poor. Wastes ~50% of bits.
|
High. Uses full range.
|
High. Activations use full range.
|
|
Weight Fit
|
Excellent. Weights are bell-shaped.
|
Good, but marginal gain.
|
Excellent.
|
|
Primary Use Case
|
Mobile/DSP (ARM SDOT)
|
Legacy (early TFLite)
|
Standard INT8 Inference
|
Compare grid constraints, correction overhead, and fit across symmetric, asymmetric, and hybrid schemes. section 2.4
For weights, the demonstrated INT8 quantizer uses codes from −127 to 127, zero-point Z_w = 0, and scale S_w = max(|w|)/127. It divides by the scale, rounds, and clamps. Excluding −128 preserves a symmetric integer range. section 2.5
For activations, the demonstrated affine quantizer extends the observed minimum and maximum to include zero, then uses codes from 0 to 255. Its scale is the real range divided by 255; the zero-point is calculated from the minimum, rounded, and clamped to the integer range. Including zero supports zero-padding and preserves ReLU sparsity. section 2.5 section 2.6
For an activation vector and a weight vector, the scaled dot product is S_x S_w [∑(q_x q_w) − Z_x∑q_w]. Setting Z_w to zero removes the correction involving ∑q_x, but the activation zero-point correction remains. It must not be omitted when Z_x is nonzero. section 2.4 section 2.5
Resolution and correction
In one ReLU comparison, symmetric quantization used codes 0–127 with scale 0.139; asymmetric quantization used 0–255 with scale 0.069, a 2.01× resolution gain. In a separate shifted-activation test, Z_x was 48: mean absolute error against FP32 was 0.0302 with correction and 1.4174 without it. section 2.5
Inspect tensor distributions and output error when choosing a grid. Activation ranges can shift between batches, so raw min/max calibration may need recalibration; percentile clipping and entropy minimization are alternatives for selecting boundaries. section 2.6
