I have a question about the FP16 issue on my hardware. As I knew, TensorRT cannot support all of devices with FP16.
Here is the document:
I wonder whether this situation is only for TensorRT?
We can see this article that he did some series experiments. The results shows 2060 is faster than 1080Ti on FP16 condition. As I checked the document above, the Compute Capability of 2060 is 7.5 and the Compute Capability of 1080Ti is 6.1 so that might be able to explain why the computing time of 1080Ti is slower. However, those experiments were tested by pytorch only, not using TensorRT.
So is there any idea about this? If I say anything wrong, please don’t hesitate to correct me.
Thank you so much!