I spent 3800$ on a new PC to use AI... Do I regret it?

Key Takeaways

I just spent more than $33,000 to go from a GTX 1080 TI to an RTX 4090 and now I'm going to do live testing to see if going from this to this was worth it I will be reacting to these tests as I do them and at the end of the video I will give my conclusion 512 512 it took like 8 seconds in my old PC so for this I'm guessing like 3 seconds would be fine and the normal with highr like this same image with high rest it took 1 minute so I would guess 30 seconds for this one would be pretty nice maybe 20 20 would be great to be honest so let's see if it works I'm going to click let's see generate waiting wait what how am I zero seconds what what what do you mean okay all right sure let's go with like the high S I can't see it my phone's on the screen oh wait it's downloading let's click again go okay the first generation is Crazy Fast 5 Seconds wow what the hell wow that's way above my expectations for this very fast I wasn't expecting much of a change in regular image generation but I definitely felt it even though it's nothing on what's to come so let's move on to sdxl these tests are a little different because in my old PC I needed to use the argument met bam in order to run as the XL properly but for this one since it doesn't really need it I just took it out which means that the other PC had a slight advantage in this case in the old one uh regular SD XL with met vam took 1.3 minutes I'm guessing with this one with 30 seconds I think that would be pretty nice let's see let's see there's no way what this is the sdxl model right yeah like 5 seconds is it's nuts for an sdxl uh model I was not expecting that like SD XL took so long on my other machine let's hit it with the highr let's yeah generate so the estimated time is 20 seconds that's crazy not going to lie okay now it should load the now this is lower okay it's slower but it's crazy fast like it's slower compared to the other one it's making a 2K image and I'm recording myself but that's crazy how much did it take 40 seconds okay so on the other PC this took uh 14 minutes damn that's that's crazy I think I could even go with bad size here I can I do this there's no way I'm doing bad size with sdxl man on the other PC I could not dream of using bad size on sdxl like no way I could just do one image that's it and if I did badge size it was without highres because highres one image with highr took 14 minutes h four images with highr just didn't even load like it it did not work it took 2 minutes to make wow man it's still saving yep that's it that's crazy man this is very nice because I started getting new possibilities that I didn't have on my old PC and of course a big jump on speed the next test I did was making an image like I did in this video right here where I basically used from control net to after detailer to latent couple to composable Laura everything it was a little bit of a fail in my part because I didn't realize that the dynamic prompts extension actually messed up my composable Laura so the results won't be great but we can still see the change in speed the regular image with no high rest and my old PC took uh 4 minutes I also could not do like batch of it I'd be happy with 2 minutes 1 minute maybe maybe 1 minute would be great seeing the the overall results of this you can see it's a lot lower than a regular stable diffusion generation 512 512 but it is going a lot faster than my old PC that's for sure all right it was 30 seconds I expected 1 minute so this is very good let's check with highest fix now I don't like using highest fix if the image is not working well but whatever let's try okay it starts immediately on my case it spent a few seconds thinking at least now it starts very fast and it isn't stopping either like it's going really really fast okay it's think the faces all right and it did the faces like immediate it was not I didn't even see what was happening here that's that's crazy okay that's great two but yeah we went from 4 minutes to 30 seconds and then we went from 15 and 1/2 minute to 1 and 1/2 minute that's a 10-fold increase that's very very nice I can't even begin to tell you how good this new pc would have been for the last video I posted because being able to experiment 10 times faster and I could even do bad size with this it means that the video not only comes out faster but I can test more stuff and be more precise with the information I give also in this case the high rest I did on my old PC was only 15 minutes but for some cases the image used the full 11 GB of bam that I had making the image generation take like 5 hours and talking about speed increments the eight bat size experiment we're going to do now is probably one of the craziest changes I've seen inside text to image all right so I'm pretty hyped out for this one I'm I'm I'm nervous for this one okay for the other one uh normal like this what we're going to do now took 40 seconds I'm expecting this to take like five like how this is going it wouldn't Shock Me if it went like 5 seconds but realistic bed probably will be more like like eight I I'd guess okay okay okay what the [ __ ] what the Yeah so five wow that's crazy man what the hell and now this is the big one okay to generate this with highr on the other PC it took 1 hour okay 1 hour one whole hour for this I'm guessing it will take less than 10 minutes if it takes less than 8 minutes I'm I'm very happy so yeah let generate let's click I'm not going to stop recording for now H yeah 8 seconds you're not taking 8 seconds to do this yeah there's no way you have to to highr BR wait wow there's no way bro there's no way like this what let didn't finish yeah this is high rest what okay so wow what the hell like so we went from 1 hour to 40 seconds what how is that much of a change by the way I can see and hear my PC the other PC was so loud I'm generating AI images with this and I'm getting no noise it's at 23° and there's no noise it's literally just quiet and this is what you would here if I were to try to record myself while on my old PC yeah about 90 times faster than on my old PC that's just nuts in my opinion from here on the rest of the tests are just utility tests instead of just testing the speed we will be testing if this new pc can pass the tests or not starting with the max resolution I can take an image to with just image to image and no scripts the idea will be to use this highres image we had import it into image to image and multiply it by two and then try to multiply it by two again so the first time I did this on my other PC uh up scanning it by two took 8 and a half minutes right so first off let's see this I'm guessing it would take like one at the speed that this this thing is going I'm not sure maybe two one maybe two I would say looks like it's going to take a minute I don't know Rick looked a lot faster than 1 minute uh 26 seconds now that's the first UPS scale on and very fast as we can see on my older PC The Next Step I Only Could multiply by 1 to 25 and that took 18 minutes on this one I'm not sure I don't know I don't know if this PC can get the image to 4K directly I have no idea I sure hope it does but I wouldn't be very upset if it didn't on my old one this would be K error uh you won't be doing this pretty much let's see this one looks like that's going to happen out of memory yeah try to allocate 36 okay so yeah we're not doing 4K directly but that's that's pretty nor oh the next best thing is bringing it to 3K let's see generate oh no no 40 GB what the hell I couldn't multiply by two either even though it's not something I'm like super sad about because that's not how you upscale anyway I did expect it to be able to make a 3K image but anyway it's not that big deal in any case I will try that too but first let's compare the times multiplying by 1.25 yep it worked all right we have our image in how much instead of 18 it took 1 minute yeah so nice speed as expected at this point and just for you I made the 4K image anyway using scripts ultimate ASD app scale tile let's see if we can on my old PC I haven't tested this because it probably would take like four four years if it did it this will probably take a while so I'm going to go beat and it took by the way uh 6 minutes uh this is not going to be like a great generation or whatever that's it 6 minutes for a 4K image that's crazy and that done with sdxl I also tried multiply an image by 1.25 this I I could not do this okay in my other PC I could not upscale with sdxl at all so maximum image I could get 2K and 35 seconds and yeah and now on to Laura training this is the part that had me the most excited about because on my old PC I could not really train or at least not at a comfortable rate so I really wanted to see what could I do with this one and also I wanted to thank Le for here as he sent me some of his tests too which means I also have lefor results on his 30/90 using the same Json and data set I will be comparing the training we did with two batches first and then I will make one with four batches my old PC took 2 hours the 3090 took 11 minutes and my new 490 spent 9 minutes and now something I wanted to touch upon is that with my old GPU I could not use either bf16 nor admw 8bit I could not use any 8bit version so now I'm going to try a well this it's the same Json file right but this time I I used four B size and then I used bf16 and amm w8p okay let's see if it works uh start training let's see if it starts fast again yeah wow okay this is way faster than before another very nice thing about this PC is that it starts straining immediately after I click instead of waiting around for like 10 minutes like my old PC used to do as far as time goes about size of four took 9 minutes for The Four's 3090 22 for his 3060 and for my 4090 you can see for yourselves okay so now it finished it's pretty crazy like 5 minutes six let's let's say 6 minutes 6 minutes for a training is very good we went from we went from 2 hours to 6 minutes that's that's very very nice and now as a surprise I just went a little crazy I think I'm going to try and train with a B five of 50 I don't know if that's even possible four okay so it's working and it's not even full bro what the hell like this could probably handle like 75 oh by the way the training of 50 batches took like 4 minutes now something I couldn't do on my old PC and that this one is opening the door too is sdxl Laura training before this with the exact same Json and data set clicking train on an sdxl Laura gave me a preview of 112 hours but now so it's been like 50 minutes since the time I hit train and now it's pretty much done oh yeah we went from 112 hours to less than one this is great that's that's awesome um I love this yes great stuff and now we only have two tests left well three depends on how you count it because the first one is just testing the luras with an XYZ plot since it's pretty much the same as just generating images normally I'm just going to give you the time for regular SD 1.5 five luras and three different weights it took less than 2 minutes compared to 15 on my older PC and on sdxl with no highest it took 2 minutes which means that now I can train and test luras without wasting hours upon hours upon hours and the last thing I want to test is very related to this topic too because the reason I say wasting hours is because I couldn't do anything else with my PC if I was generating images so the tests that are left are pretty much seec if I can use my PC while experimenting or training stuff starting with can I edit videos while using stable diffusion well the answer is it depends I tried making an XYZ plot with sdxl and highres so basically 2K images and it went well if I'm just using sdxl but when it switches from sdxl normal generation to highr fix then it starts lagging out and I can't really edit or do anything pretty much try again with just regular stable diffusion 1.5 with 512 by 512 images multiplied by two with high rest and that went totally fluid no problems at all amazing pretty much um it is still generating right yeah okay and I can see the preview I can see the preview pretty much perfectly it it doesn't stugger sometimes I either don't have anything to edit or I just want to relax bringing us to the next test can I play games while generating images I tried pretty much the same as before an XYZ plot highr fix and SD 1.5 with 512 x 512 images and then I run Elden ring and this is how that looked on my my old PC with no highres fix but this is how it looks on the new one running at 60 FPS what what that's kind of crazy not going to lie unfortunately I wasn't recording the screen properly which means you can't see the FPS but don't worry I recorded it later while testing if I could train allora and play at the same time and no it totally wasn't an excuse to retry marget after I lost to him the first time before though this is something I could not do with sdxl if I was using highest fix the same thing that happened when editing happened with Elden ring too resulting in me playing at slow motion come on no run must be said that regular sdxl with no high rest gives no problem and finally as I said before the Testo SD 1.5 training which is likely to be something I do in the near future this time recording the FPS all wa let's try to par the last time okay gg gg marget gg it's time to answer the question I had for this video was it worth it well just for context every single video I make has a lot of experimenting and generations behind it so even if I was just saving 5 minutes per generation it is such a big deal keeping in mind that I could generate like what 500 images in one video even more sometimes so for me this is 100% worth it since it opens the door to new ideas and more and better videos so I can treat it as a business expense and an investment but a more casual stable diffusion enjoyer doesn't need to generate that many images and probably doesn't care that much about speed either so before even thinking about buying a new GPU or new pc first you have to make sure that your stable diffusion is fully optimized and that you have a need for a new GPU I will of course leave some resources on how to optimize a stable diffusion in the description and if after that you still need an upgrade think about what it is that you want to be doing with with a ey and what you need for it as well as if it is likely to change in like the near future for example if you want to work with very very large images then I would recommend buying a GPU with more than 12 GB of bam the newer the better for Speed but they get also more expensive if in the other hand you just care about speed just getting a newer RTX card will help a lot mainly if you don't have an RTX at the beginning I thought about making a video testing multiple gpus thanks to the community's help and support but Shadow lak a wise on our Discord server warned me about how ineffective that would be since there are a lot of factors that I wasn't really keeping in mind therefore I will leave another Link in the description that showcases a study on which gpus are faster on stable diffusion and that it is a better done study that way you can make a budget for yourselves and what you need and there you go I just wanted to see if I can make a quick video on the topic and see if it can help someone as well as just announcing that they have a new pc so expect better videos faster and more of them I also no longer have to worry about if a video needs a l to really like work out and stuff which means that I can do more ideas and other stuff that I had in mind that I couldn't do before which is pretty nice so thank you everyone that watched this video even if it was a little weird and not really what I do in this Channel and see you on the Discord if you have any doubts and also big thanks to the people that were willing to help see you