Help on choosing a GPU for faceswap

[setxkbmap]

Hello.

Processor: Ryzen 5 5600
Motherboard: ASRock B550M-HDV
RAM: 2x16 AH4U32G32C28Y7GAA-2
PSU: HP1-J700GD-F12S

I'm building a new computer, and the last part I need is the graphics card. I am very indecisive about the choice and will use it for faceswap and similar tasks.

In my country, new gpu's:
RTX 3060 models:
Galax RTX 3060 1-Click OC 36NOL7MD1VOC 192 Bit GDDR6 12 GB > $320
MSI RTX 3060 Ventus 2X 12G OC 192 Bit GDDR6 12 GB > $335

RTX 4060 models:
Gainward RTX 4060 Python II NE64060019P1-1070V 128 Bit GDDR6 8 GB > $334
Galax RTX 4060 1-Click OC 2X V2 46NSL8MD9NXV 128 Bit GDDR6 8 GB > $336

They are being sold at these prices. The prices for used cards are very close, so I'm considering buying new. A used 2060 Super is available for $165, and if it's suitable, I can buy that.

Which graphics card would you recommend for smooth performance in faceswap tasks with this system?

[torzdf]

Go for the newer card imho (although there won't be a lot in it).

Avoid the overclocked variants. Overclocking is bad for Machine Learning, so you would need to turn off the stock overclocks anyway.

[setxkbmap]

Go for the newer card imho (although there won't be a lot in it).

Avoid the overclocked variants. Overclocking is bad for Machine Learning, so you would need to turn off the stock overclocks anyway.

Hello, thanks for your valuable comment, there is a question I want to ask to be sure:

Before I was using the RX 6600XT video card for these jobs, the program gave a warning but it still worked. Now I don't have much budget and I want to use my budget in a minimum way and get an efficient result.

You gave this https://en.wikipedia.org/wiki/CUDA#GPUs_supported in a comment you shared before and I took a look at it. my first question is, do we need to look at the CUDA SDK support score? Like RTX 2060S 7.5, 3060 8.6, 4060 8.9.

My second question is, would 2060 super or 3060 be enough for things like deep fake production?

I don't want to spend a lot since the price difference is incredibly high in my country and I haven't made money from these jobs yet, which card performance would be enough with a minimum budget?

Thanks, good forums.

[torzdf]

You gave this https://en.wikipedia.org/wiki/CUDA#GPUs_supported in a comment you shared before and I took a look at it. my first question is, do we need to look at the CUDA SDK support score? Like RTX 2060S 7.5, 3060 8.6, 4060 8.9.

Anything 10xx and above will work fine (at the time of writing)

My second question is, would 2060 super or 3060 be enough for things like deep fake production?

3060 (12GB) will be fine. I would not use it for professional work, but then I would not use any consumer GPU for professional work (you would be entering the extremely costly A6000 type range by then)

[JDBAU]

You gave this https://en.wikipedia.org/wiki/CUDA#GPUs_supported in a comment you shared before and I took a look at it. my first question is, do we need to look at the CUDA SDK support score? Like RTX 2060S 7.5, 3060 8.6, 4060 8.9.

Those values don't translate to actual performance. For example, my RTX 3060 scores 8.6, whereas my RTX 2080ti scores 7.5. But in real world Faceswap tests on GPU intensive models (Dlight 256, Villain, etc) the 2080ti is at least 3x faster.

If you want the best bang for buck my suggestion is look at second-hand RTX 70/80 cards from a previous generation, eg 2080ti, 3070, 3080, etc.

If you have to buy retail then the RTX 3060 12GB is a fine card but it's overpriced.