8GB VRAM GPU? Here are some Automatic1111 Stable Diffusion performance tests to see what works best

hey everyone I have another system here that I thought would be a great test platform for some performance testing this is a slightly older system than my other system as a core i7 this is going to be really hard to read and I don't think I can enlarge this anyway but I just wanted you to be able to see this is an 11th generation core i7 uh motherboard is a z590 and it has 64 gig of ddr4 the graphics card is a RTX 3060 TI with eight gig of RAM so I was actually going to set this up to test stable diffusion XL before I do that though I wanted to run through some performance testing and make sure I had a handle on automatic 1111 performance just kind of get my bearings so I figured while I do that I might as well take a video so that other people can benefit from some of the experimentation that I'm doing here so how I set this test up I'm going to use this positive negative prompt in each of the tests I just have this here so that you can see where I'm getting some of this information from this is the automatic 1111 GitHub wiki page there's a lot of great info here what I'm using are the optimization page and you can see what each of these things is supposed to be doing for us and then I set up a quick script that will just help me more quickly run through different setups so that we don't have to go hand editing my batch file each time I just have this all built into this already so we can just kind of iterate through all these what I'm going to do first is run no performance arguments so you can see I am using the argument Auto launch just because it makes things easier but nothing other than Auto launch and then I'm going to use just xformers Med vram low vram Med vram and xformers low vram and xformers and then opt STP attention let's get started I have my batch file here ready to rock and so here we'll just start it with no arguments okay I'm using the default 1.5 model and this is a freshly installed automatic 1111 it's got all the latest updates as of yesterday and let's put our positive prompt in let's put our negative prompt in and then I think I'm going to leave the sampler to Euler a sampling steps to 20 is fine I just want to try to make our test as consistent as possible but let's bump the batch count to four just to stress this a little more and then you can see I don't know it might be hard to see but you can see I'm using torch 2.0.1 okay so let's hit it and see how it does okay and we get a nice summary right here it might be hard to see I tried to enlarge it but we're talking 14 seconds 11. so I'm just making a quick note of that and you can you can see your iterations per second here it's just nice to have the total time taken okay so let's shut all this and start again as a matter of fact I'm going to close this whole browser so we're not having any weird caching or any other things okay let's try number two x formers make sure we up this to a batch count to four and hit generate let's see what we get okay we did save a little time using xformers the previous result was 14.11 this is 12 seconds 25. so that's not bad saved a little bit a little more than 10 percent I guess so let's shut this down and try the next one now we're trying medium vram want this to four and generate let's see full now interestingly without X formers the medium vram added quite a bit of time so without any arguments we were at 14 and about a tenth of a second with the X formers we're at 12 and a quarter of a second and now we're at 1870 so by putting the medium vram on a 8 gig 3060 card it actually added time so that's a definite no thank you now we'll try low vram well I can already tell the low vram flag is not working very well this is crawling it is worthwhile to understand you know when to use these and when not and it is worthwhile to experiment which makes me think I should definitely do this on my other system okay wow one minute 49.01 seconds holy mackerel that was awful onward we go let's kill that off oh this little guy's really happy having a good time test five which is going to be medium vram and xformers 1650 okay 1650. now that is faster than medium vram by itself not as fast as X formers by itself let's take a look at test six which is going to be low vram and X formers and let's see if that actually helps I think that's going to be pitiful I mean I think we know what's going to happen here right so okay here we are one minute 44 seconds 0.43 which was barely an improvement from when we just ran the low vram switch by itself now so far our fastest result was with just the X formers alone let's try one more test which is so far we've tried I think the the memory management medium and low vram management switches which are probably not intended for an RTX 3060 with 8 gig now we're going to try to use a option that's probably not meant for this card for the opposite reason because it's it's too low of memory which is the sdp attention that was 1239 how interesting so with this we have almost the same performance as using xformers you wouldn't use xformers and and the sdp attention so let's just try for kicks we're here to experiment right it'll take me just a second add it right under here actually I'll add it above just save this so you can see I'm using them both at once did load X formers let's see what happens yeah okay so there's no there's no performance gain it was 12 seconds 39 without xformers and it's 1240 because I think they're I think really one kind of replaces the others what we ended up with was the following looks like our fastest time by just the touch is X performance by itself looks like sdp attention is very close and then running them both at the same time is pointless it's worth mentioning if you want to try some of these experiments yourself where do you set this I suppose that would be helpful information so if you did find some of this useful you could try some of this on your own you would navigate to where your stable diffusion web UI folder is and you would want to find web ui-user dot bat you may not see the dot bat extension if depending on how your windows is set up and if you open that you'll see there are some areas where you can changes and so you can add your command line arguments here together you don't have to put any other special characters between them you can just string them together on the same line and off you go it is worth though experimenting on what works best for your whole system specifically depending on how it's built depending on what you typically keep running and all of that part of the experimenting part of the fun if you want to get the most performance you can I hope this was helpful thanks for watching this thing looks like a alien cat eating a watermelon

Channel: Renaissance Laboratories

Views: 5,552

Keywords: stable diffusion, generative AI, AI art, automatic1111, vram performance

Id: yUt065nlros

Length: 9min 49sec (589 seconds)

Published: Tue Jul 25 2023