Building an AI Art Computer
A look at the PC hardware I use for doing AI Art. Avoid overspending.
I built a PC back in September 2023 for doing AI Art. I am still using it and very happy with it. I am not thinking "I wish it was faster, had more hard drive space" or anything similar.
My job is not to build PCs or review them. In fact I build PCs relatively rarely because I am primarily a Mac user. If you are an experienced PC builder you might want to do something entirely different from me. The appeal of my build is that it is easy and hassle free. That was my concern when building this PC. I did not want to end up with something I had to waste time troubleshooting. I dreaded having to watch lots of YouTube videos to figure out how to do something properly.
My motivation for writing about this build is twofold:
- Tell you how to build a PC specifically for AI Art
- I had such a great hassle free experience that I want to share with you.
My Criteria for picking PC Components
You should know a bit about the criteria I put in to buy this computer:
- It had to generate AI Art images fast enough for me to not find it frustrating to work with.
- Reasonable size. Previously I had a large PC case that became very annoying to deal with.
- Easy to build and put together.
- Look good. Yes, I am a Mac user after all.
- Quiet, so it doesn't distract me when working
I have to stress the size part. I got a really big case earlier because I was so annoyed with how small Mac's were an how nothing would fit inside them. That was huge mistake. Big PC cases are just very impractical to handle. Just selling it becomes a problem because people will have serious issues carrying it away. Sometimes you need to move a case around the house for doing upgrades or you are relocating your work area. It doesn't really matter the reason. I simply had no idea how annoyed I would end up dealing with a big PC. It made me not want to use it.
In contrast I have been extremely happy with the size of my AI Art PC.
PC Cabinet
Fractal Design Torrent Nano White Clear Glass
This is the box you put all your hardware in. The case I picked is a compromise between space, cost and practicality. If you get a large case you get fit anything inside it so that is very nice, but then it becomes a massive hassle to move around, upgrade and place.
Get a too small case and it becomes a hassle to fit things inside and you need to buy special components that are more expensive. The Fractal Design Torrent Nano is in my view a very good compromise.
It doesn't fit a normal sized ATX motherboard. You have to use a mini-ITX board, so that is a bummer. On the other hand you can fit a normal sized power supply unit (PSU) inside it.
This case is unusual in that the power supply is place in the top rather than the bottom. However I found that this made building the PC much easier. Also the smaller mini-ITX board was much more practical to fit inside. I found it easier to handle with my hands and place inside. So for people who just want an easy build this is a win.
Motherboard
ASRock B550M-ITX/AC, AM4, ITX, B550
I picked this motherboard because it is an AM4 socket board and thus much cheaper than the newer AM5 socket boards. The socket decides what kind of microprocessor you can put in. By going for AM4 I cannot get the latest generation of Ryzen microprocessors from chipmaker AMD. On the other hand I save a lot of money which can be put into a better graphics card. For AI Art it is the graphics card which really matters so I did not see a reason to prioritize a very high performance microprocessor.
It has been over a year now since I made this choice and I have not yet regretted it. You simply don't need that powerful CPU. It is all about the graphics card.
Microprocessor
My choice in CPU wasn't particularly studied in detailed. I actually just look at what AM4 socket processor which was sold the most. This one happened to be faster than the Ryzen CPUs I had bought before which seemed plenty fast. It also had a very reasonable price. I saw no reason to blow a lot of money on CPU when the graphics card matters most.
To cool it I went with a reasonably priced fan:
Be Quiet Pure Rock 2 (155mm tall, AM5/AM4, LGA1700/115x/20xx - 150W TDP)
It is better than a stock bundled cooler and works plenty as long as you don't plan to overclock your CPU. This choice has held up well it seems. It is very quiet and it hasn't acted up and gotten louder after being on 24/7 for over a year.
I did consider buying liquid cooling, but that costs more and requires some maintenance. If you are lazy it is much better with a regular fan.
Graphics Card
ZOTAC Gaming GeForce RTX 4070 AMP AIRO, PCI Express 4.0, 12GB GDDR6X
Picking the right graphics card was tricky. I knew it had to be an Nvidia GeForce because that gives you less hassle and they generally perform better as Stable Diffusion as originally made with GeForce in mind. I had to try to balance price and performance. Obviously I would want as good performance as possible but I also don't want to blow too much money on this computer. After all I primarily use a Mac for my work and cannot rationalize spending a ton of money on a single purpose computer.
The image resolutions you can work with in Stable Diffusion is heavily dependent on the amount of memory you have on your graphics card. Thus ideally you want a 16GB card but those turned out to be very expensive. There are 8GB cards with much better performance than this card, but I picked a 12GB card to make sure I wasn't going to be significantly limited in resolutions I can work with it.
Thus when picking card there is a choice between fast generation of images vs max resolution you can work with. Resolution is a very important point to me. The typical Stable Diffusion output image is 512 by 512 pixels which isn't much. It means you will often be lacking in details. A great looking image will need higher resolution that that. Typically I get higher resolution and better details by using various upscaling techniques and image composition.
However the higher resolutions you can work with directly the easier that job will be. The GTX 4070 card can output 22 images per minute in 512x512 resolution according to Toms Hardware. However I don't know how many steps they use to get that number. Number of iterations per second for this card is around 18. Typically I use 20-30 iterations when I create an image. So it should take me a couple of seconds max to get an image.
My online service spends around 5-10 seconds doing the same. That is my reference point. Lower speed than that would be frustrating.
In terms of price relative to performance, this card is probably not the best from what I have just read. The older cards such as the GTX 2070 card would likely give very similar performance at much lower price. They are worse at gaming, but who cares if you are just running Stable Diffusion?
I wrote this assessment a year ago. How has it held up? Performance is fine. I don't feel my work is hindered or I am getting annoyed. Generally performance is better than you get with online services. I have one issue though. Working with SDXL images which are 1024 by 1024 pixels is not perfect. Generating them is fine but doing things like high resolution fixes where you try to upscale the image to higher resolutions often gets me in trouble. Thus 12 GB memory on the card is not ideal. It was a compromise. The 16 GB cards are just really expensive and 8 GB is way too small in my opinion.
An alternative I have seen people suggest is to get an older generation card such as a 3090 used. These card may not be as good with performance but they give you more memory at a reasonable price. But I would not call my RTX 4070 a deal breaker. It has worked very well for me overall.
Memory and Hard drive
For memory I got 64 GB DDR4 running at 3600MHz of type Corsair Vengeance LPX. Again not a particularly conscious choice. I got Corsair because I have bought them for years and been happy. Us Mac users would historically have least problems when getting Corsair memory. I opted to have more memory rather than faster. In general when working with AI models more memory is going to be a good thing.
I made the same choice when picking my Kingston NV2 NVMe M.2 hard drive. I could have gotten a faster 500 GB drive at about the same price as the slower 1 TB drive I opted for. Primarily I making these kinds of choices based on previous experiences. I have rarely been bothered by not having a fast enough CPU. But I have frequently been bothered by having too little memory or hard drive space.
After nearly 1.5 years of usage I will say going for the bigger drive was absolutely the right choice. Performance has no been an issue, but I already use 637 GB of my drive. I can probably clean up a lot of that, but it has allowed me to not have to care.
