integer-only inference

**Integer-Only Inference** is a **deployment strategy where the entire neural network forward pass uses integer arithmetic exclusively** — eliminating all floating-point operations to enable fast, power-efficient execution on edge devices and microcontrollers. **What Is Integer-Only Inference?** - **Mechanism**: All weights, activations, and intermediate computations use INT8 (or INT4). - **Quantization**: Scale factors are pre-computed and fused. $y = GEMM_{int}(W_{int8}, x_{int8}) cdot scale$. - **No Float**: Even non-linearities (ReLU, Softmax) are approximated with integer lookup tables. - **Frameworks**: TensorFlow Lite, ONNX Runtime, TVM. **Why It Matters** - **Microcontrollers**: ARM Cortex-M has no FPU. Integer-only is the *only* option. - **Speed**: INT8 GEMM is 2-4x faster than FP32 on GPUs (Tensor Cores). - **Power**: Integer ops consume significantly less energy than floating-point. **Integer-Only Inference** is **deployment-grade quantization** — the final step to make AI models run on the smallest, cheapest silicon.