NVIDIA RTX 3070

The RTX 3070 is a mid-range Ampere GPU with 8 GB of GDDR6 memory at 448 GB/s. In 2026, used prices around $250-300 make it one of the cheapest ways to run 7B models on CUDA. The VRAM limit means you're restricted to 7B-class models at comfortable quants.

Quick verdict

At $250-300 used, the RTX 3070 is a cheap way to experiment with local AI. 7B models run at 45+ tok/s. But 8 GB VRAM is the bare minimum in 2026 - you'll want more if you get serious.

Spec breakdown

• VRAM: 8 GB GDDR6
• Memory bandwidth: 448 GB/s (14 Gbps)
• TDP: 220 W
• CUDA cores: 5,888
• Tensor cores: 184 (3rd gen)

Real-world AI inference

Model	Tokens/sec	Source
Llama 3.1 8B Q4_K_M	~48 tok/s	Community
Mistral 7B Q4_K_M	~55 tok/s	Community
Mistral Small 3	Does not fit	VRAM limited

Honest alternatives

• Intel Arc B580 (~$250): 12 GB VRAM, better for AI
• Used RTX 3060 12GB (~$200): More VRAM, slower but more capable
• Used RTX 3080 (~$400): Faster, more VRAM