Listen, trying to force that on a single 3090 is just asking for a thermal disaster! If you have 3k and need reliability for work, you really should pivot to a multi-GPU setup to spread that VRAM load. You basically have two solid paths here. First, grab two NVIDIA GeForce RTX 3090 24GB GDDR6X cards. They are workhorses and that 48GB combined VRAM is massive for local inference. Pros are definitely the cost-to-performance ratio, though you will need a solid 1200W+ PSU to keep it stable. Alternatively, look at a NVIDIA RTX 4090 24GB GDDR6X. It is absolutely amazing for speed, but one card might still struggle with your dataset size. If you can stretch the budget, get a used workstation board with more memory. Reliability is key here, so dont cut corners on the cooling!

Just saw this thread and honestly, 3k is actually a decent budget if you ditch the consumer cards and go for used workstation gear. I was in your shoes last year and the constant crashing was driving me insane until I swapped over to a setup with more headroom. If you want something that just works, look for a used NVIDIA RTX 6000 Ada 48GB GDDR6 if you can find a deal, otherwise grab two NVIDIA RTX 3090 24GB GDDR6X units and link them up. The 48GB of VRAM on the Ada is a dream for heavy datasets and I've been super satisfied with how it handles sustained loads compared to my old setup. Using two 3090s is way cheaper though, and honestly, it works well if you have a solid power supply, like a Corsair AX1600i 1600W Titanium, to keep everything stable. Biggest tip: dont cheap out on the cooling or your case airflow. I stuck everything in a Fractal Design Meshify 2 XL and my temps dropped like 15 degrees. It's a bulky beast, but having that room to breathe is exactly what stopped the crashes for me. You really dont need to overcomplicate it, just prioritize the VRAM capacity over everything else and youll be golden for your project next week.

Honestly, a single card setup for that workload is just asking for trouble. I've been there and it is super frustrating when your workspace just freezes up mid-task. Before you dump money into new parts, I have two quick questions:

• How much VRAM are you actually using during peak processing?
• Is your current cooling solution enough to handle dual cards under heavy load? Knowing those details would make it way easier to figure out if a multi-GPU jump is actually gonna solve your crashes.