Selur's Little Message Board
[HELP] SCUNet(mlrt) running very slow - Printable Version

+- Selur's Little Message Board (https://forum.selur.net)
+-- Forum: Hybrid - Support (https://forum.selur.net/forum-1.html)
+--- Forum: Problems & Questions (https://forum.selur.net/forum-3.html)
+--- Thread: [HELP] SCUNet(mlrt) running very slow (/thread-4389.html)



SCUNet(mlrt) running very slow - andrewschen - 23.05.2026

Hybrid_dev_2026.05.20-174257
VapoursynthR73_torch_2026.02.26 / VapoursynthR76_torch_2026.05.16 both tried
vs-mlrt_2026.03.18 / vs-mlrt_2026.03.26 both tried

encode FHD 1920x1080 video, svt-av1 crf20 preset 6 temporal filter 0 with 2~4 fps

original speed (no vs filter) is up to 80+ fps
Intel 275hx / 5090 laptop / 596.36 studio driver
3A game runs fine

Will test Hybrid_dev_2026.05.22-18618 later

Is there any setting I should do with geforce driver or nvidia app?


RE: SCUNet(mlrt) running very slow - Selur - 23.05.2026

SCUNet call itself looks fine to me:
clip = vsmlrt.SCUNet(clip, model=3, backend=Backend.TRT_RTX(fp16=True,device_id=0,verbose=True,use_cuda_graph=True,num_streams=1,builder_optimization_level=3,engine_folder="D:/APPD/HybridTemp"))
The decoder call:
clip = core.lsmas.LWLibavSource(source="C:/Users/mmddffkk/Desktop/tmp/TestSrc.mkv", format="YUV420P8", stream_index=0, cache=0, prefer_hw=0)
(assuming the source can be hardware decoded) could be tweaked.
Personally I would use DGDecNV, but alternatively just enabling hardware decoding for LWLibavSource might be a good idea.
(see under "Filtering->Vapoursynth->Misc->Source")

That said, that speed seems kind of normal for SCUNet.
I get around 3-4 fps with my Geforce RTX 4080 with those settings.
Disabling CUDAGraph has no impact on the encoding for me.

Depending on the VRAM you have tiling can help. (I got 16GB VRAM)
1 tile: 3.77 fps
2 tiles: 3.84 fps (so no real change, that much can be due to general system fluctuation on a small clip)
4 tiles: 5.21 fps
8 tiles: 3.02 fps (slower than 1)
At least on my system 4 tiles seems to be the sweet spot for HD and 16GB VRAM.
As a side note: on SD sources I usually use streams 3, with tiles 1.
I also tried 4 tiles + streams 3, which gave me 5.55 fps.

=> you might want to do some testing regarding tiles&streams on your setup too Smile

Cu Selur


RE: SCUNet(mlrt) running very slow - andrewschen - 23.05.2026

Got it, misled by Gemini..... Big Grin It said SCUNet(mlrt) can be very fast....