2017 Google Pixelbook (intel i7, intel iris 615?)

[RogerDass]

Hello!

Thanks for making this amazing project!

So i'm running this with realistic vision v1.5 checkpoint. I'm getting ~75 s/it with OpenBLAS enabled.
Any idea how to speed that up significantly?

With ComfyUI on the same machine in CPU mode, I'm getting ~30 s/it and it takes 10 min to generate 512x512 image with 20 steps.

What's causing such a large performance difference?

Do you know if there's a way to get some basic OpenGL 3 acceleration for some of the tensor ops?

[Cyberhan123]

My observation is that ggml does not achieve maximum convolution optimization, my solution is to use turbo model or configure quantization, such as q8_0

[Amin456789]

use lcm lora sdv1.5 with 4 steps and taesd it be much faster

[Green-Sky]

~75 it/s

pretty sure you meant s/it

[RogerDass]

Yeah s/it...

Please keep in mind that ComfyUI is using f32 and not any lower quantization with 20 steps and is more than 2X faster.

Anyone know what tensorvision-cpu is doing that's so much faster than gglm?

[FSSRepo]

@RogerDass It's just that PyTorch implements more optimized convolution algorithms that are too complex to implement in ggml. That's why PyTorch is quite heavy; instead of reinventing the wheel, they reuse existing code to avoid unnecessary complications.