Per-Channel Quantization Techniques

• Per-channel quantization is a method that assigns unique quantization parameters to each neural network channel, improving accuracy by adapting to channel-specific statistics.
• It minimizes quantization errors by isolating outliers and managing dynamic range variability, benefiting applications like large language models and image super-resolution.
• Advanced techniques such as channel splitting and mixed-precision allocation further optimize quantization, enabling robust and efficient inference on specialized hardware.

Per-channel quantization refers to the practice of assigning separate quantization parameters to each feature channel or filter dimension in neural networks, in contrast to per-tensor or per-layer quantization, which employs a single set of parameters for the entire layer/tensor. By aligning the quantization step size (scale) and zero-point with the statistics of each channel individually, per-channel quantization adapts to channel-wise variation in signal distribution, dynamic range, and outlier structure, leading to smaller quantization errors, higher robustness to low precision, and better preservation of network accuracy or rate-distortion performance across a wide spectrum of tasks—including LLMs, image super-resolution, vision-language-action policies, and scientific compression tasks.

1. Mathematical Formulation and Per-Channel Schemes

Per-channel quantization assigns each filter/channel its own quantization step size, and optionally, zero-point. For a generic tensor (e.g., convolutional weight $$W \in \mathbb{R}^{O \times I \times K_1 \times K_2}$$), the most common strategy is symmetric, per-output-channel quantization. Let $w^{(c)}$ denote all weights associated with output channel $c$:

$s_c = \frac{\max_{i} |w^{(c)}_i|}{2^{b-1} - 1}$

$$\widehat w^{(c)}_i = \text{round}\left(\frac{w^{(c)}_i}{s_c}\right)$$

$$\text{quantized:}\quad w^{(c)}_i \approx \widehat w^{(c)}_i\, s_c$$

For activations, the same logic applies, or for data-free approaches, batch and layer statistics are used to set the per-channel scales (Yvinec et al., 2022). In asymmetric or affine quantization, per-channel zero-points $z_c$ are also computed based on the channel minimum.

Variants exist for grouping along the input channel axis, motivated by task-specific outlier structure (see per-IC quantization in (Heo et al., 2023)). Further, in autoregressive, tokenized, or VQ models, entire channel slices may be quantized as a vector to discrete codebook entries, as in channel-wise vector quantization (CVQ) (Song et al., 25 May 2026).

2. Motivation: Outliers, Dynamic Range, and Quantization Error

The rationale for per-channel quantization is rooted in empirical and theoretical observations:

• Wide Dynamic Range Variability: Channels within the same layer exhibit diverse distributions and scales. Per-layer quantization can cause “narrow” channels to be underutilized and “wide” channels to overflow or be clipped, as seen in Vision models (Oh et al., 2020), LLMs (Qin, 2024), and wireless neural receivers (Yellapragada et al., 8 Aug 2025).
• Outlier Isolation: In LLMs and super-resolution networks, outliers in a small subset of channels can inflate the shared quantization interval, drastically increasing quantization error for all non-outlier channels (Wang et al., 2024, Hong et al., 2020, Heo et al., 2023).
• Stability in Ultra-Low-Precision: Per-channel quantization makes it feasible to achieve W4A4 or int4 quantization with minimal loss, whereas per-layer/tensor quantization often collapses at such low bitwidths (Yellapragada et al., 8 Aug 2025, Hong et al., 2020).

By matching quantization granularity to intrinsic variation, per-channel quantization yields more uniform quantization error across channels and reduces worst-case approximation error.

3. Algorithms and Hardware Implementation

A typical per-channel PTQ pipeline involves:

1. Channel-wise Range Estimation: For each channel, collect statistics over calibration data (e.g., max, min, robust percentiles, or moments). Statistical approaches are common for data-free quantization (Yvinec et al., 2022), with alternatives such as clipped intervals or soft-clipping to reduce outlier influence (Hong et al., 2020).
2. Scale and Zero-point Calculation: Compute $s_c$ (and, for affine quantization, $z_c$) per channel.
3. Quantization and Folding: For activations, optionally fold the per-channel scales into the weights (as in OutlierTune (Wang et al., 2024) and MergeQuant’s QSM (Wang et al., 7 Mar 2025)) to enable hardware-efficient integer-only kernels. Similarly, per-channel input scales can be folded into batchnorm statistics (Yvinec et al., 2022).
4. Pruning/Splitting: Channels that are highly sensitive may be split into sub-channels (“channel splitting”), then low-energy channels are pruned to maintain complexity, further reducing quantization distortion (Sun et al., 2022).
5. Bitwidth Allocation and Mixed Precision: Sensitivity analysis (e.g., via output MSE, gradient statistics, action deviation, or Hessian-vector products) can guide channel-wise bitwidth (or even pruning) allocation subject to a global rate or bit budget (Xu et al., 3 Feb 2026, Shi et al., 17 Feb 2025, Hong et al., 2020).
6. Rotation/Permutation: Preconditioning weight matrices with blockwise SVD·Hadamard rotations can further equalize per-channel energy and diffuse outliers before quantization (Ω-QVLA (Wang et al., 27 May 2026)).
7. Hardware Adaptation: Power-of-two shift-based scalers approximate floating per-channel scales with integer shifts, dramatically simplifying hardware at minimal accuracy loss (Oh et al., 2020).

A summary of core algorithmic primitives (focusing on their function, not workflow) is provided below.

Approach	Core Quantization Step	Hardware/Complexity
Per-channel symmetric	$s_c = \max |w_c| / (2^{b-1} - 1)$; $w^{(c)}$0	Extra $w^{(c)}$1 scales; integer-efficient after folding
Shift-scaler (Oh et al., 2020)	$w^{(c)}$2; shift/scale per channel	Barrel shifters, small SRAM for $w^{(c)}$3
Clipping/soft-clipping	Apply robust statistics or soft truncation to ignore extreme outliers	Data-driven, but cheap once computed
OutlierTune QSM (Wang et al., 2024)	Fold scales into weights, shift bias for symmetrization	One-time update; inference overhead eliminated
Channel splitting (Sun et al., 2022)	Split sensitive channels into $w^{(c)}$4 subchannels, prune low-energy channels	Keeps model size; reduces sensitive channel error

4. Advanced Sensitivity Analysis and Mixed-Precision Allocation

Moving beyond uniform per-channel bitwidths, recent frameworks integrate channel-wise sensitivity analysis to optimize quantization under accuracy or rate-distortion constraints.

• Sensitivity Score Calculation: Sensitivity may be defined via $w^{(c)}$5 quantization loss, action-space deviation (critical in Vision-Language-Action policies) (Xu et al., 3 Feb 2026), or fully network-wise MSE (via Hessian-vector products in INR-VC video coding (Shi et al., 17 Feb 2025)).
• Global Bitwidth Optimization: Mixed-precision is posed as a constrained optimization:

$w^{(c)}$6

with $w^{(c)}$7 denoting the error at bitwidth $w^{(c)}$8. Greedy or min-heap solvers efficiently allocate bits or perform pruning (Xu et al., 3 Feb 2026).

• Per-IC vs. Per-OC Grouping: Depending on outlier structure (especially in LLMs), grouping in the input-channel (per-IC) instead of output-channel (per-OC) axis further isolates rare outliers and globally reduces error (Heo et al., 2023). Adaptive grouping can dynamically blend per-IC and per-OC based on Hessian-weighted sensitivity (AdaptDim).
• Dimensional Reconstruction: When channel-wise scale disparity remains large, splitting out large-scale channels into multiple stubs and pruning weak ones homogenizes quantization and preserves network capacity (Wang et al., 7 Mar 2025, Sun et al., 2022).

5. Task-Specific Variations and Empirical Performance

Per-channel quantization has shown significant gains over competing quantization schemes across diverse neural architectures and tasks:

• LLMs: OutlierTune and MergeQuant both demonstrate that per-channel activation (and weight) quantization can achieve W8A8 or W4A4 quantization with sub-1% accuracy drop on Llama-2-70B and OPT-66B, with up to 2× memory savings and 1.5–2× speedups (Wang et al., 2024, Wang et al., 7 Mar 2025). For especially pathological models, per-group quantization on only a small number of layers recovers all lost performance (Qin, 2024).
• Image Super-Resolution: DAQ achieves ≤0.1 dB PSNR loss at 4–6 bits per channel post-training, outperforming both static and dynamic per-layer quantization (Hong et al., 2020).
• Vision-Language-Action: Channel-wise bit allocation via action-space sensitivity in VLA models (QVLA) preserves 98–99% task success using only 30% of the original VRAM, with >1.4× speedup over LLM-derived methods (Xu et al., 3 Feb 2026).
• Wireless Receivers: 8-bit per-channel PTQ attains near-lossless BLER in high-mobility OFDM settings, while 4-bit per-channel PTQ incurs only a modest penalty; per-tensor 4-bit quantization fails completely (Yellapragada et al., 8 Aug 2025).

The following table highlights exemplary results.

Domain	Scheme	Accuracy Drop (vs. Float)	Notable Outcomes	Reference
LLM (Llama-2-70B)	MergeQuant W4A4	1.3 pts zero-shot	2.1× speedup, 3.6× smaller model	(Wang et al., 7 Mar 2025)
VLA (OpenVLA)	QVLA channel-wise INT4	≤0.5 ppt success	28% VRAM, 1.49× speedup	(Xu et al., 3 Feb 2026)
Super-Res (EDSR)	DAQ per-channel 4-bit	≲0.1 dB PSNR	Beats dynamic/static per-layer, no retraining	(Hong et al., 2020)
Wireless Rx	Per-channel int8	≤0.1 dB BLER	Matches float performance	(Yellapragada et al., 8 Aug 2025)

6. Structural Variants and Extensions

Several further refinements or alternative formulations extend the basic per-channel paradigm:

• Channel-Wise Vector Quantization (CVQ): In visual tokenization, CVQ quantizes entire channel maps (rather than spatial patches), yielding 100% codebook utilization and improved reconstruction fidelity in generative image models (Song et al., 25 May 2026).
• Per-Step/Token Scaling: For temporally-drifting activations (e.g., DiT action heads in Ω-QVLA), per-channel per-step lookup tables absorb range shifts and preserve quantization accuracy at each denoising step (Wang et al., 27 May 2026).
• Binary-Shift Scalers: Hardware-friendly variants use per-channel binary shifts (rather than arbitrary scaling factors), enabling high-throughput fixed-point inference with minimal area and power impact (Oh et al., 2020).
• Channel Splitting and Pruning: Sensitive channels are split into several “virtual” channels to reduce dynamic range, while low-energy channels are pruned to maintain network size (Sun et al., 2022).
• Data-Free Input Quantization: Leveraging BatchNorm statistics enables accurate per-channel scaling without any data or calibration, facilitating privacy-robust deployment (Yvinec et al., 2022).

7. Key Limitations and Recommendations

Despite its advantages, per-channel quantization introduces several practical challenges:

• Hardware Overheads: Requires storage (and potentially unique addressing) of $w^{(c)}$9 scale factors per layer, though the memory cost is minor compared to weights (Yellapragada et al., 8 Aug 2025). Some hardware is optimized only for per-tensor scaling.
• Calibration Cost: Accurate estimation of per-channel statistics may require a moderate number of calibration examples (e.g., 512 for OutlierTune) (Wang et al., 2024). In models without biases, symmetrization-based improvements are inapplicable.
• Low-Bit Regimes: For sub-4-bit quantization, advanced techniques such as adaptive grouping, Hessian-aware correction, or channel splitting become critical to avoid catastrophic accuracy collapse (Heo et al., 2023, Sun et al., 2022).
• Extremely Heterogeneous Channels: In certain pathological cases (e.g., LLaMA3-70B), even per-channel quantization is insufficient, necessitating per-group quantization or bi-directional smoothing to rebalance error distributions (Qin, 2024).

Recommended Guidance:

• Use per-channel quantization whenever significant inter-channel dynamic range or outlier presence exists.
• Fuse per-channel scales into pre- and post-processing kernels or weights for hardware-friendly integer inference.
• Employ channel-wise mixed-precision allocation guided by sensitivity when working under strict bit or rate budgets.
• For ultra-low bitwidths or pathological distributions, integrate splitting, pruning, or advanced axis selection informed by task-specific sensitivity (Shi et al., 17 Feb 2025, Xu et al., 3 Feb 2026, Qin, 2024).