stable diffusion performance question

[fajterini]

Hi, This is not and issue report, but I am interested what performance do you have with stable diffusion 1.4 plugin.

In my case i tested this on my laptop and old workstation and here are my results:
AMD laptop with 5800U (28W) CPU 8C/16T, W11, DDR4 16GB about 8 s/it INTEL Skylake-W2140B (120W) CPU 8C/16T, W11, DDR4 32GB about 2.9 s/it

Can AVX-512 help ? How is performance with Intel ARC 7XX GPUs or other CPUs 13gen, Zen 4?

Edit: fixed it/s to s/it (seconds per iteration)

[RhynarAI]

On A770, I get numbers between 7 and 9 it/s

[fajterini]

On A770, I get numbers between 7 and 9 it/s

Thanks @RhynarAI I was expecting that running this on GPU will be faster, but in this case CPU is enough for me :)

[RhynarAI]

The speed reported is tricky.... If its "more than one iteration per second", it states it/s... If its less than one its "seconds per iteration", so the reciprocal is given...

Your CPU numbers state "seconds per iteration" - my GPU number is "iteration per second"..

So running it on a GPU definitely is faster :)

I ran it on my CPU (mobile i7, 11th gen, comparable to your AMD) and its about the same speed as your AMD (and takes miinutes to finish) - the GPU is much faster..

From other tests with stable diffusion, a 770 seems to be in the region of a 3060 to 3070 - but I am not sure if these numbers wil stay that way because comparing CUDA vs openvino, SD 1.4 vs 2.0 etc.... can be tricky.. At this point of testing, I don´t even know if the Openvino pöugin is using FP16 or FP32 models :)

So far, the A770s performance isn´t bad for the money :)

I have AVX512 in my Intel 11 Gen CPU - it does help a little in stuff like video-encoding via x265/4, but not too much. I have no experience on how much of an impact it would make on inference engines, but my feeling tells me its only a slight boost... On youtube on Intel channel I did see a Video of running a 4. gen XEON which is pretty fast. These very new CPUs seem to have even more "AI Accelerators" on board (there is the GAN Acclelerator in recent Intel chips. but its more for 1D Inferences like speach recognition, even if calable by the compiler for SD - it should only help slightly...)

[fajterini]

Wow, thanks in this case it's perfect, i was hoping that GPU would be better than old CPU that I have. Arc 770 GPU is like 0.125 s/it comparing to 3 s/it (or 0.33 it/s) this is like 25x faster :)