File size: 515 Bytes
1933d4d b97fa3e 355785c b97fa3e 8327714 355785c 8327714 a1632f2 8327714 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 |
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>
|