First off, here's a spreadsheet for this. If you want to skip the nerd shit just look for something with a high CUDAbench score and at least 8gb of VRAM and you'll be fine.
https://docs.google.com/spreadsheets/d/1Zlv4UFiciSgmJZncCujuXKHwc4BcxbjbSBg71-SdeNk/edit#gid=0
In general, the most important thing would be to have an NVIDIA GPU with at least 8GB of VRAM for inference (generating images) -- so a 3060 Ti would work just fine. If the prices on the spreadsheet areIt would be great if the prices on the sheet were accurate, a 2080 Ti is probably cheaper anda bit more expensive, but better, though. Your CPU and regular RAM barely matter at all, but VRAM determines what you can do and CUDA score is how fast you can do it. That should allow you to do most things in a decent amount of time.
If you want to get into more advanced things like training your own models, I think the bare minimum for that is 16GB if you are using 8bit Adam. 24GB is recommended, but your cheapest general option for that is a used 3090 ($800ish I thinkgood luck finding a cheap one now on Ebay). I don't know what the price of what you're looking at is now, but if you can afford to splurge on a used 3090, you most certainly won't have to buy another GPU for quite a whileI can no longer recommend the 3090 at this time.
If you want to train models more cheaply, I have also heard of people using a Tesla M40 24GB (which is a datacenter GPU, i.e. you can't really use it for gaming), which costs a bit under $150 used on Ebay, and there's a bunch of them. The good thing about that is you can use it for model training, and it is probably the cheapest way to do so. The bad news is: 1) it's much slower than even the 3060 Ti (1/4th the CUDAbench score), 2) you can't really use workstation cards for gaming very well, 3) it's designed to be used in a server rack with blower fans and therefore only has a heatsink, you'll need to figure out a cooling solution for it. Now that's not to say this is worthless -- you can still make good use of it with Dreambooth and make custom models which can reproduce a specific style or object with decent accuracy off of a few dozen images. Those don't take long to train and for an M40 that would only take a few hours to train. But for the fastest general use you probably just want a good 8GB GPU.
NVIDIA is essentially an absolute requirement because of CUDA. NVIDIA might as well have a monopoly on machine learning computation. I think there's some stuff you can do to get some things to work which I haven't looked into, but... you really just can't.
As far as I know, each version of PCI-E from 2 to 4 doubles the bandwidth of its predecessor. I have no idea how this would translate into performance directly, and I would be a bit surprised if it was a linear relationship. The worst that could happen is it might take longer to generate results.
I am assuming this is used ones on Ebay. But I know this spreadsheet is out of date on multiple counts.
Ebay price I see a few going for ~400 (buy it now price) for the 2080 Ti, but it does appear the 3060 Ti is cheaper with a bunch under 300. I'd still go for one of the 2080 Ti's if possible.
And... it also looks like all of the 3090s got snagged. Looks like I have to update the post.
First off, here's a spreadsheet for this. If you want to skip the nerd shit just look for something with a high CUDAbench score and at least 8gb of VRAM and you'll be fine. https://docs.google.com/spreadsheets/d/1Zlv4UFiciSgmJZncCujuXKHwc4BcxbjbSBg71-SdeNk/edit#gid=0
In general, the most important thing would be to have an NVIDIA GPU with at least 8GB of VRAM for inference (generating images) -- so a 3060 Ti would work just fine.
If the prices on the spreadsheet areIt would be great if the prices on the sheet were accurate, a 2080 Ti isprobably cheaper anda bit more expensive, but better, though. Your CPU and regular RAM barely matter at all, but VRAM determines what you can do and CUDA score is how fast you can do it. That should allow you to do most things in a decent amount of time.If you want to get into more advanced things like training your own models, I think the bare minimum for that is 16GB if you are using 8bit Adam. 24GB is recommended, but your cheapest general option for that is a used 3090 (
$800ish I thinkgood luck finding a cheap one now on Ebay). I don't know what the price of what you're looking at is now, butif you can afford to splurge on a used 3090, you most certainly won't have to buy another GPU for quite a whileI can no longer recommend the 3090 at this time.If you want to train models more cheaply, I have also heard of people using a Tesla M40 24GB (which is a datacenter GPU, i.e. you can't really use it for gaming), which costs a bit under $150 used on Ebay, and there's a bunch of them. The good thing about that is you can use it for model training, and it is probably the cheapest way to do so. The bad news is: 1) it's much slower than even the 3060 Ti (1/4th the CUDAbench score), 2) you can't really use workstation cards for gaming very well, 3) it's designed to be used in a server rack with blower fans and therefore only has a heatsink, you'll need to figure out a cooling solution for it. Now that's not to say this is worthless -- you can still make good use of it with Dreambooth and make custom models which can reproduce a specific style or object with decent accuracy off of a few dozen images. Those don't take long to train and for an M40 that would only take a few hours to train. But for the fastest general use you probably just want a good 8GB GPU.
deleted by creator
NVIDIA is essentially an absolute requirement because of CUDA. NVIDIA might as well have a monopoly on machine learning computation. I think there's some stuff you can do to get some things to work which I haven't looked into, but... you really just can't.
deleted by creator
To what extent is an older PCI-E version (2 or 3, presumably, for an "ancient DDR3 i7") going to act as a bottleneck?
As far as I know, each version of PCI-E from 2 to 4 doubles the bandwidth of its predecessor. I have no idea how this would translate into performance directly, and I would be a bit surprised if it was a linear relationship. The worst that could happen is it might take longer to generate results.
the calculation is done inside the gpu right? i don't think there is much transfer on the pci-e bus with stable diffusion.
If you're using one of the low-VRAM workarounds, a slow bus is going to hurt. But you're already hurting in that situation.
Where are you finding a 2080 Ti for less than a 3060? I don't even see used 2080s for less than a brand new 3060, let alone a Ti
I am assuming this is used ones on Ebay. But I know this spreadsheet is out of date on multiple counts.
Ebay price I see a few going for ~400 (buy it now price) for the 2080 Ti, but it does appear the 3060 Ti is cheaper with a bunch under 300. I'd still go for one of the 2080 Ti's if possible.
And... it also looks like all of the 3090s got snagged. Looks like I have to update the post.