Here is a note about video cards for using with stable diffusion and some their characteristics which needed to take into account the good generation speed.
The long story short, I have GeForce GTX 1660 Super which is to slow to play with stable diffusion – the basic generation for the image 512×512 with sd1.5 model takes around 20-30 seconds, not so bad, but when you use inpaint mode it grow to couple minutes. Again, it’s enough for experiments, but requires a lot of time for the waiting. So I decided to review the market to understand what is available now (Feb 2025) for my budget. As anyone else I don’t want to pay a lot, so I decided to cut the top at 1250 euro (It’s a bit more in usd, but you can just think about it like about 1250 usd).
And the starting point should be to find the characteristics that we need to keep in mind during the research.
First of all, my personal choice is Nvidia, so I didn’t look at AMD.
Then we have
- The generation, eg 3xxx vs 4xxx
- VRAM amount 8GB – 24 GB
- VRAM frequency
- Cuda cores amount
- CPU frequency
- everything else: card size, ports amount, fans amount, led backlight, power consumption etc
And yes, the king, Mr Price.
So, I started from the characteristics of my current video card
- Name: GeForce GTX 1660 Super
- Generation: 16xx
- VRAM: 6GB GDDR6
- VRAM frequency: 14Ghz
- Cuda Cores: 1408
- CPU frequency: 1.53 Ghz
Now let’s understand what we need for the Stable Diffusion. It requires
- VRAM amount – that used to load the models and a bit to process images. The insge processing depends on the image size, so if you want to work with bugger resolution, you need more RAM.
- SD1.5 (with 2GB model file) easly eats my 6GB VRAM with the 512×512 images generation
- SDXL will eat more, because the model size is around 6GB I can assume that the required minimum will be 12Gb, keeping in mind that it trained on 1024×1024 images, and it’s recommended to use this size.
- Cuda Cores – this value will directly affect the generation speed, the time needed to generate an image.
- For example if my current card has 1408 cuda cores, then with the new card that has 4500 I can expect that the time generation will be 3x faster. So, each generation will take only 7 seconds instead of 20 or faster.
- Generation
- Each new generation appends new features, increases tech process and bus speed. Implements new technologies. So it always makes sense to choose new (higher) generation if everything else is comparable.
I checked the market and found next cards that fit my requirements
| Model, Nvidia RTX | VRAM DDR6x marked with x | Cuda cores | Cuda cores, x from 1660 | VRAM freq | GPU freq | Avg Price, eur |
| GTX 1660 Super | 6GB | 1408 | 1.00 | 14 Ghz | 1.53 Ghz | 250 |
| 3060Ti | 8GB | 4864 | 3.45 | 14 Ghz | 1.41 Ghz | 450 |
| 4060Ti | 8GB | 4352 | 3.09 | 18 Ghz | 2.31 Ghz | 500 |
| 3060 | 12GB | 3584 | 2.55 | 15 Ghz | 1.32 Ghz | 300 |
| 3080 | 12GBx | 8704 | 6.18 | 19 Ghz | 1.44 Ghz | 1000 |
| 4070 | 12GBx | 5888 | 4.18 | 21 Ghz | 1.92 Ghz | 650 |
| 4070 Super | 12GBx | 7168 | 5.09 | 21 Ghz | 1.98 Ghz | 850 |
| 4070Ti | 12GBx | 7680 | 5.46 | 21 Ghz | 2.31 Ghz | 1200 |
| 4060Ti | 16GBx | 4352 | 3.09 | 18 Ghz | 2.31 Ghz | 550 |
| 4080 | 16GBx | 9728 | 6.91 | 23 Ghz | 2.21 Ghz | 2200 |
| 3090 | 24 GB | 10496 | 7.45 | 19.5 Ghz | 1.4 Ghz | 1900 |
| 3090Ti | 24 GB | 10752 | 7.64 | 21 Ghz | 1.67 Ghz | 2800 |
to be continued
