Integer-Only Inference

A training methodology that simulates the effects of low-precision integer arithmetic during the forward and backward passes. By inserting fake quantization nodes that mimic the rounding and clipping behavior of int8 or int4 operations, the network learns to adapt its weights to quantization error. This is critical for integer-only inference because it prevents the catastrophic accuracy collapse that often occurs when a floating-point model is naively quantized post-training, especially for sensitive signal classification tasks.

A mandatory pre-deployment graph optimization for integer-only inference. During training, batch normalization layers compute a linear transformation using running mean, variance, scale, and shift parameters. Before integer deployment, these parameters are mathematically absorbed into the preceding convolutional layer's weights and biases. This eliminates the floating-point batch normalization computation at runtime, ensuring the entire model graph consists purely of integer convolutions and activations without any residual floating-point operations.

The fundamental arithmetic primitive that dominates neural network computation. A MAC consists of a multiplication followed by an addition, directly mapping to the DSP48 slices in Xilinx FPGA fabric. Integer-only inference leverages the fact that DSP slices can perform int8 MACs at significantly higher throughput and lower power than floating-point equivalents. Modern FPGAs can pack two int8 multiplications into a single DSP slice, effectively doubling the peak compute density for quantized modulation classifiers.

The gradient approximation technique that makes quantization-aware training possible. The quantization rounding function is inherently non-differentiable—its gradient is zero almost everywhere. The STE bypasses this by passing the gradient through the rounding operation unchanged during backpropagation, treating it as an identity function. This allows gradients to flow through discrete integer operations, enabling the network to learn robust representations that survive the precision loss inherent in integer-only inference.

A pre-quantization technique that addresses the dynamic range imbalance between consecutive layers. In deep networks, the weight distributions of adjacent layers can have vastly different scales. Cross-layer equalization exploits the scale-equivariance property of ReLU activations to transfer scaling factors between layers, equalizing their ranges. This minimizes the clipping error when mapping floating-point tensors to uniform integer grids, preserving the signal fidelity required for accurate modulation classification under integer-only constraints.

A dedicated hardware IP core optimized specifically for integer convolutional neural network inference. The DPU features a custom instruction set and a dedicated systolic array of MAC units designed to execute int8 operations with maximum efficiency. When deploying integer-only modulation classifiers on Xilinx FPGAs, the DPU provides a pre-verified, high-performance compute fabric that handles the scheduling, data movement, and parallel execution of quantized operations, abstracting away low-level hardware design complexity.