Mid tear PC build for that let fully try SD
Mid tear PC build for that let fully try SD
I understand that my old laptop is not enough for SD, till now I work online, but it's not so nice, I want local, so looking build PC.
Budget around +-1500 with the monitor and other parts. I was thinking about : i5-13400, rtx 3060, 1tb ssd, if need add HDD, 32 gb ddr4 ram.
But maybe some one can help me with my choices.
Is i5-13400 enough? a lot of people tell me that better i7 or i9, but I don't think is worth that +few hundreds extra
Can I after if need to add a second GPU like 4060Ti, can i run SD smoothly with 2 different GPU?
I don't have a lot of understanding about motherboards, I found info that most of popular cheaper motherboards don't support ddr4 ram and 13th gen CPU?
Any help or advice would be really nice. Want to build PC for learning and trying more things, and don't want to invest too much, because everything changes so fast
For SD, pretty much any modern CPU will be sufficient -- SD runs primarily on the GPU. All the CPU really has to do is shunt data back and forth and encode/decode images. i5-13400 will be more than sufficient -- arguably overkill, as SD will not be making use of all the cores. Much of the time the CPU will just be waiting for the GPU to finish!
32 GB main RAM is more than sufficient. The main thing you will need is not system RAM, but GPU RAM. Ideally, you want 8GB minimum, ideally 12GB, of GDDR6 GPU RAM. (6GB is, practically speaking, the absolute minimum for any decent performance at reasonably high resolution output, but will restrict you to certain workflows and workarounds, and somewhat reduced performance. You will encounter occasional issues with RAM allocation for certain workflows.) For best Stable Diffusion experience, focus your spending on the GPU, even if it means compromising on other components. (This won't help your gaming experience, but SD just needs GPU RAM + raw number crunching power on the GPU -- pretty much everything else is secondary).
Last time I checked, SD doesn't allow processing on multiple GPUs simultaneously, unless you run multiple instances (possibly there are some experimental forks out there that do this to an extent). So you can't (currently) take a single job and run it on 2x GPUs, but you can take two jobs and run each on a different CPU (i.e. running two instances of SD). Multi-GPU support is being worked on, though, so maybe by the time you save up for a beefier GPU, who knows...? But to be honest, if you can go straight for the 4060 if you can, even if it means lowering the specs of your other components. But make sure you have a decent reliable PSU rated for whatever GPU + other components you have, with headroom.
Agreed, and multiple gpu are worthless for Stable Diffusion... I run a 12gb 3060 and a 4gb 1650, and I can only use one for SD. I can pick at launch so I have two bat files, one for each, and if I plan to also run a LLM then SD goes to the 1650 as LLM are stupidly hungry for vram. Plus a thing I learned when I started to mess around with distributed loads is that you really want similar performance all around: it's a lot of fun seeing an upscale going 100% load on everything, then fast gpu goes idle while slow one finishes the same amount of work, then another spike and idle... and sadly even without using the power of multiple gpus, stable diffusion can't use the extra vram, not even as fast swap. So yeah, unless things change multiple gpu for stable diffusion are a waste of budget.