Estimate the parameter and memory compression ratio when quantizing a model, for example from FP32 down to INT8, to plan deployment.
FP32→INT8 约 4× 压缩,需校准保精度。
| Item | Formula / Note | Source / Standard |
|---|---|---|
| Bit width ratio | the ratio is source bit width over target bit width such as thirty-two over eight | IEEE 754-2019 floating point |
| Memory estimate | parameter count times bit width over eight gives the memory footprint in bytes | Jacob et al. 2018, integer quantization |
| Accuracy loss | low bit quantization introduces error and needs calibration plus layer sensitivity analysis | ISO/IEC TR 24028:2020 AI trustworthiness |