Benchmark FPS Much Higher Than Export FPS on Radeon 9070 XT

Just got the 9070 XT here. I am trying to upscale a 720p source to 1080p using Proteus.

This is the benchmark result:

Topaz Video  v1.0.1
System Information
OS: Windows v11.24
CPU: Intel(R) Core(TM) i5-14500  95.755 GB
GPU: AMD Radeon RX 9070 XT  15.922 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1280x720
Benchmark Results
Artemis		1X: 	53.72 fps 	2X: 	31.50 fps 	4X: 	08.56 fps 	
Iris		1X: 	77.64 fps 	2X: 	40.93 fps 	4X: 	09.81 fps 	
Proteus		1X: 	64.32 fps 	2X: 	35.28 fps 	4X: 	10.01 fps 	
Gaia		1X: 	38.82 fps 	2X: 	25.78 fps 	4X: 	09.75 fps 	
Nyx		1X: 	24.69 fps 	2X: 	21.50 fps 	
Nyx Fast		1X: 	59.19 fps 	
Nyx XL		1X: 	69.04 fps 	
Rhea		4X: 	08.50 fps 	
RXL		4X: 	08.29 fps 	
Hyperion HDR		1X: 	38.90 fps 	
4X Slowmo		Apollo: 	81.30 fps 	APFast: 	155.31 fps 	Chronos: 	42.58 fps 	CHFast: 	67.62 fps 	
16X Slowmo		Aion: 	68.01 fps 	

However actual export is only at around 10-15 fps. I have tried with many different videos. All yield the same slow export.

I have also tried all the different models and they all have much slower export fps as compared to the benchmark.

Running Rhea caused the whole system to crash and reboot.

anything I can do?
logsForSupport.zip (90.8 KB)

Edit: I also noticed that during the benchmark, gpu load would reach up to 85-90% for each model but it would go up to 30% in the actual export.

For both cases cpu load is around 50%

Upscaling by whole numbers: 1X, 2X, & 4X is faster than going from 720p to 1080p.

Was the Original 640x360, or did you get it online somewhere at 1280x720? I ask because Videos from YouTube give you multiple resolutions options to download, but it doesn’t always mean the quality is better just because the resolution is higher…. if the Auto is 360p download that one and upscale it to 540p 1st then upscale that export x2 which will be 1080p and most likely faster overall.

As for the benchmark vs real render query:

Try upscaling X2 export instead of 1080p… Yes I know you want 1080p, but the benchmarks are showing you results of 1X, 2X, and 4X. You do not need to do a full render… just let it run for a few minutes and then look at the fps it is running at etc.

I just tried with another 1280x720 video.
Exporting with Proteus at original resolution yields 18fps.
And exporting at 2X gets 12fps.

I have a RTX 4070 Ti Super and would be willing to test your actual video with the same setting to see if I get similar results compared to my benchmark run. You can upload the video anonymously using Gofile - Cloud Storage Made Simple (no need to create an account)… then reply to this message with the download link you will be provided.

The benchmark only measures model inference performance. Actual export speed also depends on the input/output codec, disk speed, memory, and CPU scheduling.

On systems like the i5-14500, the CPU can become the bottleneck during export.

1 Like

Make sure ‘Recover detail’ is off (default is on). This setting is rarely useful (at least for me) and uses a lot of CPU/RAM performance (not covered in benchmark).

Also, ‘Dynamic Parameters’ also reduces performance because it analyzes movie frames constantly for parameter adjustment (not covered in benchmark).

As encoder use AMD GPU h264/h265 if it’s quality is enough for u. GPU encoder is faster than CPU based encoding.