AI Upscaling - DirectML vs CUDA vs TensorRT
Could anyone post examples how much faster CUDA / TensorRT would be for Polaris, Sirius and Rigel? (face impovement not activated)
Unfortunately I have an AMD GPU. I cannot compare, but I would like to know how much time I would save after upgrading to a specific NVIDIA card. So please don't forget to post which card you are using.
-
@Mirela Radu: Can somebody ban this user?
-
After I upgraded from an AMD to a NVIDIA (RXT 4060) I can answer myself.
I did several tests. The default settings of Audials will always only use DirectML which is in generally the slowest way to upscale. At least for my card.
Test 1: Anime 480p (16:9), Upscale to 1080p, 4106 frames, Antaris, predicted time: 9:10 min
DirectML: 9:03 min
CUDA: 7:02 min
TensorRT: 7:55 min (incl. 1:40 min initialisation)
Test 2: same anime from test 1, but 8951 frames, predicted time: 16:20 min
CUDA: 15:29 min
TensorRT: 14:16 min (less than 10 sec initialisation)
---------------
Test 3: film from digital cam 480p (4:3), Upscale to 1080p, 1188 frames, Sirius
DirectML: 10:15 min
CUDA: 7:00 min (incl. 1 min initialisation)
TensorRT: 4:54 min (incl. 2 min initialisation)
Test 4: film from digital cam 480p (4:3), Upscale to 1080p, 3420 frames, Sirius, predicted time: 30:00 min
DirectML: did not try
CUDA: 18:26 min (incl. 1 min initialisation)
TensorRT: 8:11 min (less than 10 sec initialisation)
I do not know when TensorRT needs that long initialization. Depending on the length of the film CUDA could be the better choice. For Antaris TensorRT is for my card a little bit better than CUDA. For Sirius it is the clear winner.
The good question is why Audials is ignoring CUDA & TensorRT by default.
Please sign in to leave a comment.
Comments
4 comments