Integer-only arithmetic

Integer-only arithmetic eliminates floating-point overhead by executing neural networks and signal processing using 8-bit or 16-bit fixed-point math.

This approach replaces power-hungry floating-point units (FPUs) with streamlined integer logic to boost performance on edge hardware like ARM Cortex-M microcontrollers and Hexagon DSPs. By mapping 32-bit float tensors to 8-bit integers (int8) via scale factors and zero-points, developers achieve 4x memory reduction and massive throughput gains. It is the gold standard for deploying low-latency AI on hardware lacking native float support, ensuring consistent deterministic results across diverse silicon architectures.

https://www.tensorflow.org/lite/performance/quantization_spec