12GB VRAM vs 16GB+ VRAM difference / 4070 Ti Super / 50 Series chat

[magicAUS]

I'm going to update my video card as for the life of me can't get my 12GB VRAM AMD 6700 XT to work with this outside of low VRAM mode, along with a third party software continuously crashing with the current AMD GPU.

Is there going to be much difference between 12GB VRAM and 16GB VRAM using Fooocus? I.e. RTX 4070 Super vs RTX 4070 Ti Super.

[Rogala]

I have not found a comparison between 4070 and 4070 ti super, I myself want to buy 4060 ti 16GB, I came across an article and video with its testing, it clearly shows how a large number of VRAM helps to speed up the process.

At the end there is a link to the test 4070 ti super

Second link to a Japanese review between 4060 ti and 4070 super, nothing is clear but the graphs are interesting.

[Irmagon]

RTX 4070 Ti Super now is best choise speed/price for SD, short of buying used 3090 or 4090

[Rogala]

Nobody argues with this, but is it worth it while waiting for series 5***, if you are involved in SD in a financial environment, then yes, but if just for the sake of interest.

In my opinion, 4060 ti 16 GB is what allows you to wait a year.

[uberneko]

Hello -
Just stumbled across that benchmark -
https://www.tomshardware.com/pc-components/gpus/stable-diffusion-benchmarks

I also would have thought 16Go instead of 8Go would be better for the 4060 Ti - but here on both tables the 16Go one... performs lower ?!?

[mashb1t]

Fyi this benchmark is from December 15, 2023 and does not include SDXL.

We haven't tested SDXL, yet, mostly because the memory demands and getting it running properly tend to be even higher than 768x768 image generation.

In general you can future-proof your setup by rather getting a GPU with more VRAM than the one with less while also keeping your requirements on hardware demands (gaming, image generation, web browsing, etc.) in mind.

[uberneko]

Thanks !
Ok sorry I'm quite new in AI... so I should push for higher VRAM GPU then -
That's becoming a difficult exercise for me as I also would like to be able to play around still with Blender (I'm much more advanced here) - the requirements AI / 3D do not seem to be totally the same.

[Rogala]

I don't want to create a new topic, all rumors and news about 5000 series.

Key takeaways from this information.

Many people who follow this 5000 topic assume that the 4000 series will soon become more expensive, since its reserves are artificially low and the price will rise.
Nvidia is preparing people to buy the new series.

[magicAUS]

I don't want to create a new topic, all rumors and news about 5000 series.

Many people who follow this 5000 topic assume that the 4000 series will soon become more expensive, since its reserves are artificially low and the price will rise. Nvidia is preparing people to buy the new series.

Updated this thread title to include 50 Series chat. :)

[kalle07]

I find that 12GB is fine if you only use Ai, be it LLM(GPT) local or text2img.
if you train loras -> 16GB is very useful
i have the 4060Ti (16GB)

[magicAUS]

I ended up purchasing a 4070 Ti Super which should arrive late next week.

My MOBO is PCIe 4 (looking like v5 needed for 50 Series), and CPU would be a bottleneck for 50 Series as well.

So I'll keep this setup for 1.5-2 years and review thereafter.

[barepixels]

get a used 3090. it's the best bang for the bux

[magicAUS]

get a used 3090. it's the best bang for the bux

Depends where people are located. In Australia, a used 3090 usually costs more than a new 4070 Ti Super, sometimes $400-1,000 more.

It also draws a lot more heat and power, so need to consider that when we're talking "bang for buck". And a high amount of system RAM is needed to get the most out of the 24GB VRAM.

I'm loving the 4070 TiS so far. I'm not using complicated SDXL workflows. The generating and saving time using Fooocus realistic higher res is 9.09 seconds, which I'm content with. Fooocus only seems to use ~7-8.5GB of the 16GB VRAM atm.

[magicAUS]

Not much difference:

Upscale 2x of 1472 x 704 images w/ --always-gpu & NVIDIA control panel setting changed:
15.4GB VRAM utilised
image 1: Generating and saving time: 50.30 seconds
image 2: Generating and saving time: 50.27 seconds

Upscale 2x of 1472 x 704 images w/ Settings reverted to normal all-around:
8.8GB VRAM utilised
image 1: Generating and saving time: 51.29 seconds
image 2: Generating and saving time: 51.07 seconds

[Rogala]

You should ask @mashb1t why there is no acceleration, if everything goes through VRAM and GPU it should be faster.

I don't know much about the intricacies of the job FOOOCUS

[magicAUS]

@mashb1t , thoughts? Other PC specs FYI:

i5-12400F
32GB DDR4 RAM
Win 11 Pro 64-bit 23H2
NVMe SSD

[mashb1t]

@magicAUS if usage of either VRAM or RAM doesn't make a difference it's your GPU compute power bottlenecking the process.

[magicAUS]

@mashb1t I have an 850W PSU (MSI MAG A850GF), and GPU Power is 260-286 W during rendering (100% power, memory clock ~10,500 MHz). GPU OC software is set to 100% power target. Set vram state to: HIGH_VRAM is displayed.