Oh man, I feel u — SFF + AI is basically “VRAM Tetris” with heat as the timer lol. For your situation, I’d prioritize VRAM + efficiency over peak FPS. For local LLMs and SD, running out of VRAM is the hard stop way before “raw shader speed” matters.

Here’s what I’d do (years of trying to make quiet small boxes not melt…):
- **Best all-around pick:** NVIDIA GeForce RTX 4070 Ti SUPER 16GB — 16GB is the sweet spot for comfy SD and decent-sized LLM quant runs, and it’s usually way easier to undervolt than people think. In a 2-slot card, it’s not always available, but if you can find a true 2-slot model it’s kinda amazing.
- **Quieter / easier SFF pick:** NVIDIA GeForce RTX 4070 SUPER 12GB — less VRAM, yeah, but it’s efficient and tends to behave in tight cases. If you mostly do 7B–13B quants and SD at sane resolutions, it’s fine.
- **If you can snag a deal used:** NVIDIA GeForce RTX 3090 24GB — VRAM monster, but tbh it’s a heat/power diva in SFF. Only recommend if you’re ok undervolting + accepting more fan noise.

Performance hit vs VRAM? Honestly, going from “faster GPU” to “enough VRAM to not offload” usually feels faster in real use. Undervolt + set a power limit and you’ll dodge the jet engine problem. good luck, cheers

+1 to reply #2 — BIG warning: dont chase “faster core” and end up VRAM-starved; you’ll hit OOM and everything tanks. In SFF, also avoid triple-fan monsters + high-TDP cards: heat/noise spikes hard, and you’ll throttle anyway.

Pro tip: before u buy, run ur exact LLM/SD targets through a VRAM estimator + benchmark lists. For SD/LLMs in SFF, it’s basically VRAM > peak speed.

Option A: NVIDIA GeForce RTX 4060 Ti 16GB — coolest/quietest-ish in 2-slot, 16GB is actually usable for 7B/13B + SD, but perf per $ is… meh.

Option B: NVIDIA GeForce RTX 4070 SUPER 12GB — way faster, but 12GB is the hard wall (bigger SD models / higher res / bigger context). Also many cards are chunky.

Option C: used NVIDIA GeForce RTX 3090 24GB — VRAM king, but heat/noise in SFF is rough.

Resources: check Hugging Face model cards for VRAM notes, and use TechPowerUp GPU database + local LLM VRAM calculators (like “LLM VRAM calculator” search) to sanity-check. gl!

> I have room for a 2-slot card and I’m trying to keep power/heat under control Yeah honestly, as a fellow DIYer just starting out, I think building it yourself is the only way to go. Pre-built "AI workstations" are usually massive and way overpriced, so doing a custom SFF build is way cooler. I’ve been researching the Inno3D GeForce RTX 4070 Ti SUPER Twin X2 because it’s one of the rare 16GB cards that’s actually a true 2-slot design. It’s a bit pricey but seems perfect for a DIY setup where you're trying to squeeze every bit of VRAM into a tiny case. I'm kinda nervous about the heat in such a small box tho—is undervolting something a beginner can handle? I’ve heard it’s basically the best way to keep these tiny builds from sounding like a jet engine, but I don't wanna mess up my hardware lol. Going the DIY route definitely feels like the move if you wanna customize your fan curves and keep things quiet, especially compared to some loud professional blower cards. Does it really make a huge difference if I just leave it stock for a bit?

> VRAM seems like the biggest limiter, but I’m not sure how much performance I’d be giving up if I prioritize VRAM over raw GPU speed. tbh i’m still pretty new to the AI scene but i’ve been researching this like crazy... i think the consensus is basically that VRAM is king no matter what. but i’m wondering about the whole brand war thing? like, i’ve heard people say the "green team" is the only way to go because of the software, but then i read somewhere that the "red team" is catching up and usually offers way more memory for cheaper? im not 100% sure if the setup for non-nvidia cards is gonna be a total nightmare for a beginner though. someone told me the drivers can be a bit of a mess for local LLMs unless youre a linux wizard or something lol. basically just feels like a trade-off between "it just works" and "more ram for less money." has anyone looked into the market lately to see if the price gap is even worth it anymore? its all sooooo expensive anyway...

Works great for me