Best Graphics Card for Training Stable Diffusion - Which GPU TO choose - nVidia Intel or AMD?

in this video we shall discourse of graphics cards and their uses for stable diffusion I've always been fascinated by graphics cards and their mysterious Powers they are like magic boxes that can transform the images on the screen into something alive something real the combination of different graphics cards and stable diffusion this is Magic this technique enables us to produce lifelike images which are of great utility in the fields of computer Graphics staple diffusion is employed in various Arts and Sciences but it Demands a graphics card of great power and skill which can perform the complex calculations and rendering necessary having said this I shall compare several graphics cards of different brands and qualities and judge their speed their excellence and their price and also how big and hot they run we start off by considering memory now there are three architectures of memory to consider one is Unified memory which is found in Apple devices the second is shared memory if you have a Windows PC with a graphics card discrete graphics card you are probably using shared memory and then there is the discrete graphics called memory this is memory that you find when you buy the graphics card on the box you're gonna have a memory score that memory number is n't of discrete graphics card memory that is what we're really interested in when it comes to staple diffusion that's probably the most important number now if we go back a few years ago there were a bunch of gpus known as the GTX gpus they're still around but these ones are kind of fully out of favor for high-end uses like stable diffusion they have a limited amount of memory and that limits their utility when it comes to producing Graphics in machine learning there are three manufacturers to consider when it comes to discrete graphics card and there is another company which you consider which is Apple Apple used what's known as unified memory and unified memory in Apple today the recently announced Mac Pro can provide up to 192 gigabytes of unified memory that's memory that is shared between the graphics and also the system and it can be used by either the ecosystem for Apple isn't so good when it comes to machine learning so most of the recommendations will be either Intel AMD or Nvidia as we shall find out the recommendations are all going to be in video there are advantages to going with Nvidia there are disadvantages and the main disadvantage is cost First recommendation is going to be the RTX 3060 this is an old graphics card from several years ago it is still available as new and the ones that you're looking at on the screen here they're currently at the lowest price that they've been ever and they've just dropped in that price fairly recently I suppose that's because they've been some recent releases of new cards from both Nvidia and AMD now both Nvidia and AMD said in May of this year that they were going to release several cards over the summer and we've just started to see those releases the RTX 3060 provides 12 gigabytes of RAM and as such it is a pretty decent offering in terms of memory now the memory is important when it comes to resolution with stable diffusion the initial inputs were generally about 512 by 512 pixels if you want to go beyond that you need to consider more memory that initial amount of 512 512 pixels it's very small however is considered to be to process that you would need about eight gigabytes of memory as a minimum when we look at the 12 gigabytes of the RTX 3060 we're looking at a reasonable offer however the power the overall power of the RTX 3060 is not that impressive it is beginning to show its age unfortunately and it doesn't have the same power as the more recent gpus the next GPU I'm going to recommend is the RTX 4060 TI this is the most recent release coming from Nvidia there is another version which is the 4060 TI 16 gigabyte version that's the one that is my key recommendation it's fairly expensive at 500 it's not going to be out until July but if you get a chance to get your hands on that one that is the one that I would get there aren't many excellent quality gpus with 16 gigabytes of RAM at 500. the other cousin is the 4060 TI eight gigabyte version that one is okay it doesn't have a huge number of advantages over the previous 30 series but it does run cooler and quieter the eight gigabytes should be enough for entry-level training now I've done two videos on graphics cards for stable diffusion by stability Ai and those two videos are going to recommend more or less the same gpus I've taken the opportunity with this particular video to express a design that we have more competition that will maybe next year will have apple competing against these other companies and maybe giving the windows platform a run for its money at the moment we are really focusing just on Windows that's where stable diffusion works at its best 4060ti will give you the ability to work with models that are fairly small and not very demanding the next recommendation is going to be the RTX 4070 12 gigabyte this is the upgrade to the RTX 3070 it is a lot more powerful about 20 more powerful than the 3070 on top of that and importantly it has 12 gigabytes of memory gdd r6x memory and that means it's particularly powerful you can get this bad boy at around the same price that the RTX 3070 was at when the RTX 3070 launched consequently I would say is a decent upgrade at about approximately 20 percent extra performance for about the same price a pretty decent card in this family is the RTX 4070 Arrow overclock 12 gigabytes this one comes from gigabyte and gigabyte is a company I tend to like now if you want to know more about the companies that are involved in graphics cards in Nvidia graphics cards I did a video last year which covered the entire almost the entire range of brand that produce these cards I'll link you to that video I would strongly recommend watching that video if you're buying a Nvidia GPU there is another card here which is the RTX 4070 from Asus this bad boy actually comes in at just 600 and I think part of the reason that Asus at the moment are selling their products somewhat cheaper than they used to is maybe because of some really bad publicity that have got over the last few weeks but like I say if you want to know more about the individual Brands follow the video in the link it will give you an idea of who you are dealing with now the 4070 is a powerful GPU with 12 gigabytes of RAM you can work with slightly higher resolutions the most powerful GPU both gaming and professional is the RTX 4090 this guy comes with 24 gigabytes of vram however consider this is only three times more than the vram in the entry level 4060 TI this is an issue because you really want to increase the amount of memory to take into account increases in resolution so for the time being I would recommend focusing on the single card that will give you the performance that you need if it is the 1490 bear in mind that you will probably be purchasing the most powerful GPU that we're gonna have for the next maybe the next 12 months so it's a good long-term investment you might go for something like this Zotac uh try to get the five-year warranty they put an extended warranty which requires registration so this is something that I talk about in the video that I mentioned before now this is a chart showing the price of the RTX 3070 in the United States and as you can see the price that kind of fluctuation is insane and it's something that we almost never see in consumer electronics but because of the cryptocurrency mining craze we saw extremely high prices for gpus for that reason I'm going to recommend purchasing the most powerful GPU that you can comfortably afford and how onto that for the long run in case you get some more of this nonsense now if your system can run the RTX 4090 it can definitely also run 4080. this is significantly less expensive but it's also much less powerful at 16 gigabytes of vram there are certain things that you can do with the 4090 that you simply can't do with 16 gigabytes and that's especially going to be the case if you're working with very very large base models if you're doing professional work there are some gpus like the RTX a6000 there's a new one called Ada Lovelace generation this one has 48 gigabytes of gddr6 error correcting memory so it's ideal for the kind of tasks where you need to get precision and it's it's best suited for working with the workstation but you can use it inside any any desktop that's capable of providing the the power that you need there's an older generation the a6000 not the average generation a the generation one but the previous one that one can go up to 96 gigabytes if you have two of those working in tandem that ability has disappeared with the latest one for some reason but these ones will be the ones you'd want to use if you have very very large data sets you want to work with base models you're not just adapting a pre-existing model uh and also finally we've got the Grace Hopper architecture which is going to be a computer that Nvidia is going to come up with later on that is for your data centers that's probably going to be tens of thousands of dollars and we also have other lots of other gpus for the data centers the h100 a100 tens of thousands of dollars and apparently they need tens of thousands of those to work with uh models like chat GPT guys that's going to be it I'll have links to some of these cards in the description hope you found that useful I'll see you in the next video oh good what about a late night [Music]

Channel: Pixovert

Views: 11,876

Keywords: Stable Diffusion choose amd nvidia apple intel, stability ai stability.ai, ai graphics, computing requirements machine learning, gpu recommendations ai, creating ai models, ai art, nvidia rtx a6000, top 5 gpus Stable Diffusion, deep learning gpu, nvidia Stable Diffusion rtx, GPU training Stable Diffusion, building a stability ai pc computer, ai vram, machine learning requirements, graphics cards for Stable Diffusion, ai hardware best, rtx 4060ti 4090 4080 3060

Id: 04Rg27smgQU

Length: 11min 39sec (699 seconds)

Published: Thu Jun 08 2023