| Resnet50 Test Quantization for accelerated int8-int8 compute for Apple Neural Engine | |
| Resnet From Code:<br> | |
| Int8ANE.ipynb - test notebook to create <br> | |
| <br> | |
| CoreMLTools 8.01b, M4 iPad Pro 16GB, iPad OS18.1 beta<br> | |
| <br> | |
| LUT 4-bit FP16 1.03ms<br> | |
| A4W8 0.58 ms<br> | |
| LUT 4-bit (A8W8) 0.92ms - no Accell!<br> | |
| Tourch Model: | |
| 8-bit<br> | |
| resnet50-LUT8-iOS17.mlpackage 1.08 ms<br> | |
| resnet50-W8A8-iOS17.mlpackage 0.81 ms<br> | |
| 4-bit<br> | |
| resnet50-LUT4-iOS17.mlpackage 0.93 ms<br> | |
| resnet50-W4A8-iOS18.mlpackage 0.68 ms<br> | |