i have build a homelab from used components:
z840 hp workstation
Intel(R) Xeon(R) CPU E5-2683 v4 @ 2.10GHz 2.10 GHz (2 processors)
128 GB RAM
Nvidia Quadro K2200 - 4GB
Tesla p40 24GB
i use Automatic1111 and ComfyUI and i'm not sure if my performance is the best or something is missing, so here is my results on AUtomatic1111 with these Commanline: -opt-sdp-attention --upcast-sampling --api
Prompt
a girl standing on a mountain
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 2573751789, Size: 512x512, Model hash: 1fe6c7ec54, Model: juggernautXL_version6Rundiffusion, Clip skip: 2, Version: v1.6.0-2-g4afaaf8a
(full body:1. 2), (Smokexxx:1. 4), 1girl with really wild hair, (wavy multicolor hair:1. 2), (two long braids:1. 2), (long wild hair:1. 2), (multicolored hair: 1. 9), wearing a long skirt, high contrast, epic cinematic, soft studio lights, rim light, absurd, amazing, funny, itricate, hyper detailed, ultra realistic, soft colors, 8K
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 2341168399, Size: 512x512, Model hash: 1fe6c7ec54, Model: juggernautXL_version6Rundiffusion, Clip skip: 2, Version: v1.6.0-2-g4afaaf8a
Time taken: 26.5 sec. 1.27s/it
A: 7.55 GB, R: 8.42 GB, Sys: 10.2/22.5 GB (45.5%)
for ComfyUI, it took ages on the first generation, i think because it try to load the model and these are the results on the same prompts with the default workfolow:
Prompt
a girl standing on a mountain
1.25it/s
Prompt
(full body:1. 2), (Smokexxx:1. 4), 1girl with really wild hair, (wavy multicolor hair:1. 2), (two long braids:1. 2), (long wild hair:1. 2), (multicolored hair: 1. 9), wearing a long skirt, high contrast, epic cinematic, soft studio lights, rim light, absurd, amazing, funny, itricate, hyper detailed, ultra realistic, soft colors, 8K
1.65it/s
please note the difference between first prompt it/s and the secont prompt sec/it on Automatic1111.
is this the best performance i can get, or is there is some tweaks that can make performance better ?
I've been thinking of getting one of these because a 3090 is too much $$$. It varies depending on exactly what the task is, but it's about half a 3090.