GPUDojo - GPU Prices by VRAM

For local AI inference, VRAM determines what model sizes you can run. This table ranks GPUs by cost per GB of VRAM — the metric that matters for running models locally. New to used GPUs? Read our guide.

Model	VRAM	BW	TDP	Arch	Cond	Tier	p/s	t/s	Cap	Speed	Price	$/GB	Link
Loading...

Articles

Scores

Cap (Capacity): VRAM² × √(t/s)

What can you run? Rewards VRAM heavily — bigger models = more capable.

Speed: 30% p/s + 70% t/s

How fast does it feel? Normalized 0-100. Rewards actual inference speed.

Speed Columns

p/s (prefill) — Prompt processing speed. Limited by TFLOPS. Matters for RAG and long prompts.

t/s (generation) — Text output speed. Limited by memory bandwidth. Your "reading speed."

VRAM Tiers

8GB — 7-8B models (Llama 3.1 8B, Mistral 7B)
12GB — 14B models (Qwen 2.5 14B)
16GB — 14-30B, MoE models (Mixtral 8x7B)
24GB — 30B+ models (Qwen 2.5 32B)
48GB+ — 70B models (Llama 3.1 70B Q4)

Budget Tips

Best $/GB: Tesla P40 — 24GB for ~$180. No display output, needs active cooling.

Best all-around: RTX 3090 — 24GB, fast prefill AND generation.

Budget consumer: Used RTX 3060 12GB or Arc A770 16GB.

Sister Site

DiskDojo.com — US storage prices ranked by $/TB.