VRAM required

#52
by ehristoforu - opened

Tell me how much minimum VRAM is needed for stable operation of the model?

24G. with everything loaded.

24G. with everything loaded.

I'm wondering how much VRAM is required to boot the 4GB model.

well... the 4GB file is the same model. it has no text encoders included.

It depends on many factors I guess... right now, using sd3_medium_incl_clips_t5xxlfp8.safetensors with my 4070S, with no extras and a simple prompt, for a 1024x1024 20 steps with Euler Sampler, nvidia-smi shows a 5.2Gb. With SDXL models it shows about 7.4Gb. (Iv'e just reached 5.8Gb producing 1080P images, more often going down to 5.6Gb)

That's pretty dope.

Yes, for the medium one, even 6GB card will work fine.

the OP asked about "stable operation of the model" which i assumed meant in an active inference API situation, where 24G is needed to keep all of the models (TEs, VAE, Transformer) loaded.

What will make it a winner will be community's production of add-ons. It's been a long time we have hoped for a SD1.5 killer, dreaming of SDXL quality but with SD1.5 lightness. SD3 2B brings even more quality, and is lighter than SDXL for sure, but it's still about 30% heavier and slower than SD1.5. It's not a huge gap to jump from 3 to 5Gb, and for most users having 8Gb GPUs it's not such an issue than loading SDXL 7.5Gb. It sounds close enough to me, and with the benefit of a way better base than 1.5, but I don't know if the 1.5 crowd will agree and jump in the wagon, and SD3 needs them to get a great variety of add-ons, it's a crucial point. As for sure it is not this base model that will be enough for most of our needs.

We'll see what Pony folks will do, it also depends a lot on the ease (or complexity) of training. That's an unknown too. But if they achieve a great Pony (or Wildcard etc) first try, SD3 may become the new tiny beast in town.

Sign up or log in to comment