Whats the absolute best GPU I can get right now for stable diffusion without breaking the bank? Im starting a new project next month where I need to churn out hundreds of images for a local indie game and my current rig is just dying. I read a bunch of threads saying the RTX 3060 12GB is the budget king, but then I see others arguing that I really need 16GB of VRAM or I will regret it later. Im based in the UK and really cant justify dropping two grand on a 4090, but Im so anxious about buying something thatll be obsolete or slow in six months. Should I just save up more or is there a middle ground?
Honestly, skip the 3060 if you are doing actual production work for a game. I have been running local models for years now and in my experience, VRAM is the only metric that matters once you hit a certain baseline. If you run out of memory, you are stuck with system RAM swap speeds and your generation time tanks from seconds to minutes. I tried the 12GB route back in the day and it just wasnt enough for training LoRAs or running complex workflows with ControlNet. If you want a sweet spot that isnt a 4090, here is what I would look for:
> VRAM is the only metric that matters once you hit a certain baseline. @Reply #1 - good point! VRAM is king for sure. If you want a solid middle ground, check out the NVIDIA GeForce RTX 4060 Ti 16GB VRAM. It is super efficient and that 16GB buffer makes a massive difference for local workflows. If you can find a used NVIDIA GeForce RTX 3090 24GB VRAM, that is honestly the ultimate move for production work.
^ This. Also, if you are doing production work for a game, speed is gonna matter just as much as the capacity. Definitely look into getting an extra drive like a Samsung 990 Pro 2TB NVMe SSD to keep your model files and cache snappy. Loading times for massive checkpoints get old real fast when you are churning through hundreds of assets. Good luck with the indie project!