Choosing a Graphics Card for Generative AI

If you’re picking a graphics card for generative AI, your best bet is to get an nVidia card. nVidia is a popular brand for Generative AI because they include chips and software that helps accelerate AI work.

If you have a non-nVidia graphics card, it might still work, though. Even though Fooocus is optimized for nVidia cards, it has special instructions for running on AMD. Intel says that their ARC GPUs are capable of running Stable Diffusion, and offers this video on installing ComfyUI to work on Intel ARC GPUs.

VRAM (the onboard memory on your graphics card) is essential for AI work. According to this chart, the easy-to-use interface Fooocus can run on a minimum of 4 GB to 8 GB of VRAM, depending on which graphics card you're using. But working with only 4 GB of VRAM can severely limit what models and functions you can use. For most purposes, the Automatic1111 interface, ComfyUI interface, and Stable Swarm all recommend at least 8 GB of VRAM, on a newer nVidia card. (Looking beyond the graphics card, for a minimum system, make sure your computer has a reasonably new CPU and at least 16 GB of RAM, as well.)

The newest and most powerful downloadable model, Flux from Black Forest Labs, takes full advantage of 24GB of VRAM, and slows down if you have less. Slimmed-down versions of this large model have been developed to work with less VRAM, though, although the quality of the smaller models isn't entirely as good as the full-size model they originally released.

Ideally, if you were buying a graphics card with AI in mind, you'd want as much VRAM as available. Systems with more VRAM can run faster in many situations, because they are less likely to be slowed down by swapping models out to RAM and back into VRAM. It will also help you move into creating animation, higher resolution imagery, training your own models, and using the latest and most powerful models.  If you're trying to get a graphics card with as much VRAM as possible (in a consumer-oriented card), then for desktops the nVidia GeForce RTX 3090 (used) or 4090 (current) have 24 GB of VRAM, and for laptops the GeForce RTX 3080 Ti has 16 GB of VRAM. 

On a Windows computer, search for 'dxdiag' into the desktop search box or Start menu. When it opens, click the 'Display' tab. (It may say Display 1 or Display 2 if you have multiple cards.) This will tell the exact model of your graphics card(s), and also will tell you the amount of VRAM. Note that the amount of VRAM is listed in MB, so divide by 1024 to get the VRAM in GB.

Some AI image generators work on Macintoshes. You can use DiffusionBee, which is made for Macs, or install Automatic1111 by following these instructions. The performance of your system might be slower than a Windows of Linux system with a more optimal graphics card. If local generation of images doesn't work well enough for you, all of the online image generators reviews on this site can be used from a Macintosh.

Technically, it is possible to use your CPU only. Automatic1111 has instructions and warnings about how to do this.

As a practical matter, though, this process can be painful, even if you get it working. If you don't have a suitable graphics card, exploring online image generators might be much more rewarding for you.