Best GPUs for Stable Cascade and Diffusion - 2024

hey guys this is Kevin for pixel.com and this video I'm going to be making recommendations for graphics cards for stable Cascade and stable diffusion so this is a stable Cascade I've been using it now for a couple of weeks and it's an amazing model from stability AI it's similar in some ways to stable diffusion but in other ways it's very different and the requirements are much more challenging than the requirements for stable diffusion at least the current versions of stable diffusion so we're going to be making recommendations for Cascade and also for diffusion but bear in mind that the requirements for Cascade can go well beyond the recommendations that we are going to make depending on which versions of uh stable Cascade you decide to use now I've got a course for stable Cascade and not everyone has got the memo that stable Cascade is different from stable diffusion but I definitely have and what I've decided to do is to treat Cascade as a completely different software from diffusion so one of the things you'll find on the course is a requirement for 20 GB of vram this originally comes from stability AI who suggested 20 GB of vram that requirement I've been able to work down and down through testing to 12 gabt of vram for the GeForce cards for the gaming cards and that's using the specific installation that is used in this particular course so it is possible to work down from about 20 GB to lower than that if you're using a specific installation which is good at memory management and you're also limiting yourself to workflows that can work with lower amounts of memory now as I say it's just been a couple of weeks since uh Cascade was launched and in their release memo stability AI mentioned that the expected vram requirements were 20 GB but they also said these can be further lowered to the further lowered by using the smaller variants now I've not done all the testing on these smaller variants they do mention that the smaller variants may also decrease the final output quality right now I'm using the large files that are provided by uh by stability Ai and these larger files would be the 3.6 billion parameter files for stage C and the 1.5 billion parameter uh models for Stage B the the approach that I'm taking is that I want this one to be higher quality than sdxl so I'm not really interested in using less powerful models if the quality level is going to go back down to stable diffusion levels some of the community have they already started producing their own thirdparty models and with stable Cascade some of the models can get very large we've got one here at 34 gabt for the stage C now with the stable Cascade files from hugging face from stability AI the largest ones are 14.4 GB and these ones found that they can actually work with uh the workflows that we have for uh 12 GB uh graphics card managed to make them work with an 8 GB card as well but that's something not I wouldn't necessarily recommend if you want to know a little bit more about that you might want to subscribe and I'll be talking a little bit more about what I did there to kind of compress the requirements for uh vram we're going to be making recommendations these are going to be in the description now Amazon as I've mentioned before do have a layaway scheme whereby if you want to you can purchase in installments that might be useful if you find a very good price but you don't want have the money to pay pay for that all at once right now or alternatively if you want to wait a bit of time to get the very best price if you're selling your previous card on eBay or whatever there is a new version of stable diffusion stable diffusion 3 which is coming along the models that they're working with range from 800 million to 8 billion parameters now that's even more in some cases than the parameters for stable Cascade so bear that in mind so the first recommendation is going to be for the RTX 360 uh 12 GB variant now I've noticed that a lot of these have been described as revision 2 not really sure what has been revised but with this one you do get 12 GB of vram which is going to go a long way to allow you to process uh some of these larger models now the RTX 360 is quite an old card relatively speaking it's a 30 series card this particular one this revision to has been around for just about a month or so so they they're still in demand they're still fairly popular and we're still seeing new versions of these even even if we're not entirely sure what the new versions involve now I did take I didn't have time to to examine what revisions have been made but if you want to you can go and check on the websites for some of these cars as to what the differen is between the revisions normally it's something which is slightly improved which uh demands a new revision we have got some uh cars from Maxon this is a new company from China mainland China their card is slightly more expensive but it's a three fan solution and it's got pretty decent feedback on Amazon that's what really Drew my attention to this I don't know that much about Maxon but I do kind of like what they've got here in terms of what they're offering three fans quite a lot of bling bling if you like a little bit of bling bling they've got a lot of bling bling and they're also using purple on their box which I think is an indication that they're probably good people now if you don't want to spend too much money on the RTX 360 12 GB that's understandable it is an old card and I'm going to whisper it it's kind of a little bit it's a little bit too expensive for what you get if you can afford something a little bit more I would definitely recommend doing so and it will be the 16 gab variant of the RT x460 this has far many more Cuda cores and it also has 4 GB more vram that's going to make a lot of difference with stable diffusion and we're talking about a $100 difference I would definitely be looking if I can to spending that little bit extra on getting this uh more powerful card more Cuda cores more memory you you really can't go wrong now at $429.99 this Zotac card is almost certainly going to be the cheapest you'll find anywhere so that's going to be the base price there are more expensive ones and I did come across this MSI gaming card now this one is a three fan variant which means it will fit in fewer cases because because however typically the three fan uh cars are have a they they have a higher clock speed they can run a little bit faster and that's because they've got better Cooling and quite often the cooling is also a little bit quieter than the two fan ones another one that I know noticed was this MSI gaming X slim card once again that excellent 16 GB of vram somewhat faster speed we have it quite a lot more in terms of the cost however you are getting um quite a lot of stuff that you're not getting at the entry level and the reason I'm really pointing these ones out is because the next card we're going to recommend is a lot more expensive than the 4060 TI 16 GB now the marketing here I really like they've got the exact size of the card and on Amazon sometimes you can't really tell whether the measurements are for the of the card or the box and you do need to know these measurements obviously to see if it will fit inside your case and the most important thing obviously is the bling bling and we got we got a little bit of bling bling here I can't quite tell if that's purple but if it is the these are probably good people and we also have the eight pin connectors it's always nice to know cuz you don't know what power supply you're going to need and it's nice to know here we can use pretty much any power supply which can provide sufficient power uh I like the marketing from uh MSI they they do tend to show you exactly what you need here we have the frozer technology which means that when the uh temperatures are cool you don't have any fan speeds so you have no noise and then when things get a little bit hotter the fans do start to ramp up and you're not going to find a lot of marke a lot of reviews so just go over to the website it will tell you a little bit about what you need to know about the card the specifications this one it requ it's recommended 500 W power supply uses up 165 wats that's going to fit inside most PCS and for something as powerful as this that's a really nice uh option to have the next recommendation is the is the RTX 460 TI super so this is one of the ones that came out in January and you're seeing here a small selection some of these guys are at MSRP that's the PNY one 799 MSI have got one at 870 and I would probably at this sort of price range I would probably stretch to some of these slightly more expensive ones rather than trying to squeeze myself into the into the entry level ones I have noticed that the ones that reviewed very well in January they're not available in the in the shops and that may be due to very large demand for those ones but the 4070 TI super has much much more powerful Cuda cores and it has a larger memory bandwidth as well so you do get much more performance than the 4060 TI 16 GB but quite honestly the 4060 TI 16 GB is a really good starting point now moving over to UK side we've got this MSI RTX 480 super gaming X Trio this one is around the MSRP for these cards and it's a powerful card much more P powerful than the 4060 TI 16 GB somewhat more powerful than the 470 TI super and it's quite a lot less expensive than the 4080 so we've lost about $200 or $300 2 or300 on the cost of the 4080 I would say that's a win for everyone and I can see why these are very popular right now in the United States you'll find it very difficult to get them around $11,000 you'll find them at around 1,400 which is ridiculous but there seems to be a lot of demand in the United States in the UK things are a little bit more reasonable and at this sort of price you really can't go wrong I think this is like3 or $400 less 304 less than we were paying for these cards just a few months ago and the big boss the rgx 490 at 24 GB this one is going to be much better at handling the really large models from Cascade uh it's going to handle sdxl with relative ease but like I say some of the Cascade files some of the Cascade models can get very large and this is not necessarily going to be ideal for working with those very large files and on top of that if you want to do training the RTX 490 although it is very powerful I think for training there may be better cards out there so the RTX 490 is still the top recommendation if you want a gaming card that can do stable diffusion it is the best out there and the prices are now reasonable they were kind of ridiculous for about 3 or 4 months if you want to know more about other options I will link to the RTX 490 and other options video from back in November December and that covers the professional cards it covers cards with more than 24 GB of vram and it covers apple as well so I'll link to that one you may want to see that one if you want the more powerful cards for training or if you just want more powerful cards for inference as well guys that's going to be it for this one I'll have the links in the description I will see you in the next [Music] video

Channel: Pixovert

Views: 1,307

Keywords: Stable Cascade, Stable Diffusion, Nvidia RTX, GPU Recommendations, AI Modeling, Graphics Card Guide, VRAM Requirements, 2024 GPU Review, AI Processing, High-Resolution AI, Machine Learning Hardware, RTX 3060, RTX 4060 Ti, RTX 4090

Id: eLNx5dpvDVM

Length: 13min 22sec (802 seconds)

Published: Sat Feb 24 2024