Here is a note about video cards for use with Stable Diffusion and some of their characteristics that need to be taken into account for good generation speed.
Long story short, I have a GeForce GTX 1660 Super, which is too slow for playing around with Stable Diffusion. Generating a basic 512×512 image with the SD1.5 model takes around 20–30 seconds – not too bad. However, when using inpaint mode, the time increases to several minutes. It’s fine for experimentation but requires a lot of waiting. So, I decided to review the market to understand what is currently available (as of February 2025) within my budget. Like most people, I don’t want to overspend, so I set my budget limit at 1,250 euros (roughly the same in USD).
The first step was to determine the key characteristics to consider when researching video cards.
First of all, I personally prefer Nvidia, so I didn’t look at AMD.
The key factors include:
- Generation (e.g., 3xxx vs. 4xxx)
- VRAM amount (8GB – 24GB)
- VRAM frequency
- CUDA core count
- GPU clock speed
- Other factors: card size, number of ports, number of fans, LED backlighting, power consumption, etc.
And, of course, the king – price.
My current video card
Here are the specs of my current GPU:
- Name: GeForce GTX 1660 Super
- Generation: 16xx
- VRAM: 6GB GDDR6
- VRAM frequency: 14GHz
- CUDA Cores: 1408
- GPU clock speed: 1.53GHz
What do we need for Stable Diffusion?
- VRAM – This is used to load models and process images. The image processing VRAM usage depends on resolution – higher resolutions require more VRAM.
- SD1.5 (with a 2GB model file) easily maxes out my 6GB VRAM when generating 512×512 images.
- SDXL needs even more, since the model itself is around 6GB. The recommended minimum for SDXL is likely 12GB, given that it was trained on 1024×1024 images.
- CUDA Cores – More cores directly affect generation speed.
- For example, if my current card has 1,408 CUDA cores and a new one has 4,500, I can expect roughly 3× faster generation – cutting the time from 20 seconds to about 7 seconds (or even faster).
- Generation – Each new GPU generation introduces better technology, higher memory bandwidth, and improved efficiency.
- If all other specs are comparable, choosing a newer generation is always preferable.
GPUs that fit my requirements
I checked amazon.it and found the following cards that match my filter. I also included some top and popular cards like 3090, just for comparison:
| 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 |
As you can see, the most balanced card is the 4070 Super, which will be approximately 5× faster than my 1660 based on the number of CUDA cores alone.
