Here I’m using an old GTX 1050TI, Intel processor, latest drivers, and the update 3.1 make my renders almost 4x slower… It made my normal routine render drops from 0.4 fps to 0.1 fps.
Returned the 3.0.12 version.
1 Like
I run 2 1080ti’s with a 5900x…3.1 is wayyyy faster than 3.0.
There is something up with 3.1 though as I said in another post, without turning on “low power mode” it gets a bit reckless. Ive seen it dedicate so many resources to a 1st instance that a 2nd instance ran at over 2spf which is outrageously slow for my machine. It works far better with low power mode on and I’ve never had to use that before.
You might want to try tweaking things a bit because 3.1 can definitely run much faster than 3.0
GTX cards took a performance hit with the 3.1 release as they don’t support TensorRT. I’ve posted about it already in the 3.1 release thread. Pretty sure the devs are aware. This specifically affects Chronos models.
That’s not really true, unless a 1080ti is different from other GTX cards. 3.1 is generally about 2x as fast on my setup and I’m happy about that. But like I said, I have to keep “low power” mode on to constrain whatever resources it’s trying to use, but with low power on, it is still 2x faster than 3.0 in most cases. So hopefully they don’t “overfix” it.
Try it with Chronos Fast only and post the results between 3.1 and 3.0.12.
I don’t have 3.0 installed anymore but I always use chronos fast alongside whatever upscaling model I’m using unless it’s DDV …
Since it is simple enough to do and I currently have a 1080P video loaded. I’m getting 0.22spf @ 29.97 → 59.94 fps using Chonos only on 3.1. I don’t think that is much faster than 3.0 but I know it’s not slower.
Your results could be due to the dual card setup. I can definitively say I get 4x slower performance with 3.1 vs 3.0.12 using Chronos Fast, and there have been multiple reports from people with GTX cards seeing similar performance issues.
That’s possible, Ive seen people say similar things, it’s only been out for a day. I do know there are issues, like if I turn off low power mode my machine is gonna crash for no apparent reason.
What kind of stress is your GTX card under while it’s running 4x slower? I know 3.1 isn’t doing a very good job of constraining itself. My machine is 100% watercooled, both cards are MSI Red Dragons and I have 2 big radiators so I have no heat problems, except getting the heat out of the room lol
Maybe the cards are running out of VRAM…a 1080ti has 11gb. I’m not sure how the VRAM is split up with 2 GPUs but effectively my machine has 22gb VRAM. I know with the 3090’s and greater they did try to utilize higher amounts of VRAM. Someone should post up a test video of whatever your having problems with
I think this is a big part of the equation. The beta prior to the 3.1 release would actually crash my entire PC if I had max memory set to 100%. Troubleshooting directly with a dev, we lowered the memory usage setting and that helped but performance was still bad.
So probably VRAM related as well as TensorRT not being supported by the Pascal/GTX cards are 2 factors.
Yeah, I’m pretty sure this is a VRAM problem. I have a 1080 with 8GB. Using 0.8 VRAM limit and with “Low Power Mode” off it goes straight up to 100% VRAM usage and becomes very slow. I guess the VRAM limit is not entirely exact, cause with 10% usage for just other Windows stuff, I’d imagine 0.8 should be fine.
Turning on “Low Power Mode” fixed things for me, now running chf+alq only uses around 70% VRAM. It looks like enabling “Low Power Mode” sets the “instances” parameter in the ffmpeg filters to 0, instead of the default 1. I don’t really get what this “instances” is supposed to do, the help says it’s the “number of extra model instances to use on device”. I haven’t noticed any performance difference when increasing it in 3.0.12, it just uses a bit more VRAM.
So yeah, it looks like the new version just needs a lot more memory. In 3.0.12 the same chf+alq setup only used around 30%. On the bright side, when it works, 3.1 is actually a lot faster.
Version 3.1.5 fixed my problem, and made my routine renders a little faster.
