Performance Guide

[aifartist]

This is less a discussion than a top level post I'll keep correcting and enhancing based on feedback and instead of burying the good stuff in endless replies. This is a work in progress and a placeholder. The goal is to provide easy to understand facts on performance related issues in this one stop shop. This is version minus 1! No complains please. At least not till I get to version 0. :-)

What do I mean by performance?

• P1: The fastest possible one image at a time generation of 512x512 images; and
• P2: The fastest throughput in terms of images per second using batching or other techniques.
• P3: The messy stuff that doesn't easily fit into the above.

I will discuss them all.

Bite sized facts(These will be expanding upon):

• There are three performance factors for both a CPU and a GPU.
  -- Hardware Speed/Efficiency/Capability:
  -- Algorithm Efficiency: When xformers came out it provided a big boost in perf.
  -- Software Efficiency: A great algorithm can be ruined by a bad programmer that can't write efficient code. Also compiler optimizations.
• A expensive fast GPU with a cheap slow CPU is a waste of money. Definitely true for P1.
• Software options which some think always help, instead hurt in some setups. It is complicated.
• Memory footprint has to be taken into consideration. If something is a bit faster but takes 2X the memory it won't help everyone.

Time to sleep.

[kopyl]

Any way to just faster launch web ui instead?)

[ash0080]

Any way to just faster launch web ui instead?)

PCIE 4.0 7000MB/s + DDR5 RAM = <4s 🤣

I'm not kidding, the biggest time consumption is loading the model, so lazy loading is also possible. However, I believe that a high-speed SSD is necessary for increasingly large models.

[kybercore]

Any way to just faster launch web ui instead?)

PCIE 4.0 7000MB/s + DDR5 RAM = <4s 🤣

I'm not kidding, the biggest time consumption is loading the model, so lazy loading is also possible. However, I believe that a high-speed SSD is necessary for increasingly large models.

Your GPU also needs to support that PCIe version for the record to maximize the speed