-
-
Notifications
You must be signed in to change notification settings - Fork 310
inference time on GPU
Yihui He 何宜晖 edited this page Sep 21, 2017
·
2 revisions
inference time on GPU depends on several factors
- GPU (GTX TITAN X in paper)
- Batch size (less than 32 is inefficient)
- CUDNN (5 in paper)
- CUDA (8 in paper)
- Caffe
- OS (ubuntu 16 in paper)
The results may varies a bit in each run.
the performance of VGG-16 3C 4X hasn't been tested in our paper. Mr. 장대석 kindly tested it on GTX1080 (linux):
| 3C 4X | Original | |
|---|---|---|
| FW | 3.6ms | 6.9ms |
| BW | 4.8ms | 5.8ms |
| Mem | 6698 MiB | 5156 MiB |