| Layer | Feature | FP16 | INT8 |
|---|---|---|---|
| Convolution | ✔ | ✔ | |
| Dilation | ✔ | ✔ | |
| Winograd | ✔ | Not implemented in SW | |
| Deconvolution | ✔ | ✔ | |
| With padding | Not implemented in SW | Not implemented in SW | |
| Winograd | Not implemented in SW | Not implemented in SW | |
| Fully Connected | ✔ | ✔ | |
| Winograd | Not implemented in SW | Not implemented in SW | |
| Group Convolution | ✔ | Not implemented in SW | |
| Winograd | ✔ | Not implemented in SW | |
| Pooling | ✔ | ✔ | |
| Max | ✔ | ✔ | |
| Min | ✔ | ✔ | |
| Avg | ✔ | ✔ | |
| Inclusive padding | ✔ | ✔ | |
| Exclusive padding | Not supported in HW | Not supported in HW | |
| Activation | |||
| Bias | ✔ | ✔ | |
| BatchNorm | ✔ | ✔ | |
| Scale | ✔ | ✔ | |
| Sigmoid | ✔ | Not implemented in SW | |
| Tanh | ✔ | Not implemented in SW | |
| EltWise SUM | ✔ | ✔ | |
| EltWise SUB | Not supported in HW | Not supported in HW | |
| EltWise MIN | ✔ | Not implemented in SW | |
| EltWise MAX | ✔ | Not implemented in SW | |
| LRN | ✔ | Not implemented in SW |
| Framework | Status |
|---|---|
| Caffe | ✔ |
| ONNX | Future |
| Network | Configuration | fp16 | int8 |
|---|---|---|---|
| MNIST | nv_full,nv_large,nv_small | Verified | Verified |
| ResNet-18 | nv_full,nv_large,nv_small | Verified | Verified |
| ResNet-50 | nv_full,nv_large,nv_small | Verified | Verified |
- Not supported in HW
- Dilation with Winograd
- EltWise SUB
- Pooling and convolution layers where pad size is greater than kernel size
- Not implemented in SW
- Deconvolution with strides > 32
- Deconvolution with input/output padding