FixQuant

Quantization research often stops at model accuracy, while hardware implementation starts from a separate set of fixed-point and interface decisions. FixQuant closes that gap. It trains 8-bit fixed-point CNNs, checks their integer arithmetic in software before any hardware is involved, and ships them as versioned, checksum-verified packages that an FPGA accelerator can compile directly.

FixQuant is the model-preparation component of TileCNN, my reusable model-to-FPGA framework.

What it does

• Trains in fixed point. Quantization-aware training with Trained Quantization Thresholds (TQT) learns power-of-two scales for 8-bit weights and activations. It uses MSE calibration and optional cross-layer equalization.
• Validates before hardware. An integer digital twin reproduces the accelerator’s convolution, requantization, pooling, and residual arithmetic bit for bit. Deployed accuracy is therefore known before a bitstream is built.
• Ships reproducible releases. Every model is an immutable, versioned release with recorded accuracy, quality gates, and SHA-256-pinned artifacts. Nothing is fetched implicitly, and nothing is transcribed by hand.

The exported package contains the network graph, integer weights and biases, validation inputs, and integer reference outputs. A manifest records the release identity, preprocessing, and a checksum for every file.

Model zoo

Top-1 accuracy (%) on the full ImageNet-1k validation set (50,000 images). FP32 is the pretrained floating-point model, QAT is the trained fixed-point model, and Integer twin is the bit-exact integer model that the FPGA executes.

Δ is the end-to-end change from FP32 to the integer twin. MobileNetV2 uses cross-layer equalization.

Publication

The workflow was published at the 28th IEEE International Symposium on Design and Diagnostics of Electronic Circuits and Systems (DDECS 2025).