3060 12gb cannot utilize all it's Vram to its full potential because the core is very weak and gutless..my 4070ti super smokes the 3060 on a bad day..i own both..3060 12gb is a great starter card..but this 40's series is really a step up..in gaming and a.i...i wish i could have afforded a 4080 super though...its a monster...but i still love this 4070ti super.
I am suspecting that an option may be slowing down things a little because I would expect a 4070ti supper to be a bit faster than that. To give an idea I recorded 512px benchmark on a 7900xt for a comparison. ua-cam.com/video/Vlh-nQv4IQg/v-deo.html
Check my Amd RX7800xt (win\zluda). On plain job is only 20-30% slower, but totally sucks on big batch and the upscaler. Last job with R-ESRGAN 4x+ upscaler took 3.5 minutes, but resize from extra tab was done in a few seconds without any issues. Strange, but ok.
I'm getting this card tomorrow: RTX 4070TI SUPER WF3 OC 16GB D6 GIGABYTE My other components: 5900x 32 RAM Some SSD's Do you think will run fast for SDXL at 768 x 1024 with *ADETAILER* on to get a good face, will mostly be using HYPER / PONY models that don't need too many steps. Do you think i will get good speeds?
It largely depends on the model and any other things you have loaded at the time into SD. For instance I use Instant ID quite a bit which means two other elements need to be loaded through Control Net which takes overall memory usage to around 9.5 GB. SD is loading models and parts of models in and out of GPU memory and system RAM all the time with my 3070. This really slows the process down. Maybe 100 seconds to 2 minutes to create a 1024x1024 image with Control Net options added. The 4070TI Super does those same tasks is 50 seconds to a minutes. Half the time However if I just use the model with no extras (Juggernaut X usually) it run a almost as fast as the 4070 Ti Super. 85% of the speed maybe? Close enough for me not to care too much. Much of the time saved seems to be memory management. If the model fits into your GPU RAM with a bit left over, you are usually good to go. It's the extras that kill low VRAM cards. But the more I use SD the more extras I use. I'd love a really capable AI card with 40GB of VRAM for under $1200. But I doubt we'll be seeing that any time soon.
Could you please share what is the max. resolution you can render in txt2img without going OOM? I mean without any tricks like NeverOOM, just pure txt2img? Also how long does it take to render 4k on this card? It took all day on 1060🤷♂
My pc with a 4070 Ti Super is arriving tomorrow and I can't wait to utilize it for this
Thanks
I'm planning to buy 4070TI super at the moment. Thank you for the review.
3060 12gb cannot utilize all it's Vram to its full potential because the core is very weak and gutless..my 4070ti super smokes the 3060 on a bad day..i own both..3060 12gb is a great starter card..but this 40's series is really a step up..in gaming and a.i...i wish i could have afforded a 4080 super though...its a monster...but i still love this 4070ti super.
Using "SDXL Lightning" and "TensorRT" can potentially double the speed.
i set the same ones for meters but use Sampling steps at 15 and CFG Scale 2 to activate lightning
45 seconds with a 12gb 3060
I am suspecting that an option may be slowing down things a little because I would expect a 4070ti supper to be a bit faster than that. To give an idea I recorded 512px benchmark on a 7900xt for a comparison. ua-cam.com/video/Vlh-nQv4IQg/v-deo.html
Hmm
I think ill stick with my 4070 ti 12gb till the next wave of 24gb-32gb cards release
Check my Amd RX7800xt (win\zluda).
On plain job is only 20-30% slower, but totally sucks on big batch and the upscaler. Last job with R-ESRGAN 4x+ upscaler took 3.5 minutes, but resize from extra tab was done in a few seconds without any issues. Strange, but ok.
I'm getting this card tomorrow:
RTX 4070TI SUPER WF3 OC 16GB D6 GIGABYTE
My other components:
5900x
32 RAM
Some SSD's
Do you think will run fast for SDXL at 768 x 1024 with *ADETAILER* on to get a good face, will mostly be using HYPER / PONY models that don't need too many steps.
Do you think i will get good speeds?
hey how its going with it?
u have 32gb ram is it enought or should go for 64gb?
i have the same speed with 3070
It largely depends on the model and any other things you have loaded at the time into SD. For instance I use Instant ID quite a bit which means two other elements need to be loaded through Control Net which takes overall memory usage to around 9.5 GB. SD is loading models and parts of models in and out of GPU memory and system RAM all the time with my 3070. This really slows the process down. Maybe 100 seconds to 2 minutes to create a 1024x1024 image with Control Net options added.
The 4070TI Super does those same tasks is 50 seconds to a minutes. Half the time
However if I just use the model with no extras (Juggernaut X usually) it run a almost as fast as the 4070 Ti Super. 85% of the speed maybe? Close enough for me not to care too much. Much of the time saved seems to be memory management. If the model fits into your GPU RAM with a bit left over, you are usually good to go. It's the extras that kill low VRAM cards. But the more I use SD the more extras I use.
I'd love a really capable AI card with 40GB of VRAM for under $1200. But I doubt we'll be seeing that any time soon.
Could you please share what is the max. resolution you can render in txt2img without going OOM? I mean without any tricks like NeverOOM, just pure txt2img? Also how long does it take to render 4k on this card? It took all day on 1060🤷♂
can it generate faces clearly?
You'd need extensions such as adetailer, or use the hires-fix feature. By default faces will look deformed at low resolutions
wow ty for video, i try to buy rtx 4070 ti super, look fasterrr