| Neural Net | Top-1% 32/32 (W/A) | Top-1% 2≤W/A≤8 | Top-1% 1/1 (W/A) |
|---|---|---|---|
| AlexNet | 60.0 | 58.2 | 57.0 |
| MobileNet-v2 | 72 | — | 54.4 |
| ResNet-18 | 69.1 | 67.0† | 64.8 |
| VGG16 | 71.1 | 69.1 | 68.9 |
| Inception-v3 | 78 | 76.4‡ | — |
† Employs ternary data representation for both weight and activation values.
‡ Employs 4-bit weights and 8-bit activations. The weights of the first and last layers are 8-bit values.