Operator coverage analysis counts which graph operators execute in INT8 versus FP32. It helps explain why quantization may compress a model without speeding inference: operators left in FP32 do not benefit from integer execution. section 6.5
Interpret coverage by checking which operators the quantization format marks as quantized:
- In QDQ-format ONNX graphs, inputs from DequantizeLinear nodes identify operators that can have the QDQ pattern fused into an INT8 kernel. section 6.5
- In operator-replacement graphs, distinct nodes such as MatMulInteger or DynamicQuantizeLinear indicate quantized operators. section 6.5
Coverage explains different outcomes
ResNet-18 ORT Dynamic quantizes 4% of compute operations: its MatMul/Gemm classification head is quantized, while its convolution layers remain FP32. ORT Static reaches 100% coverage. BERT-base ORT Dynamic reaches 28%, leaving operators such as LayerNormalization and Softmax in FP32.
