LLM VRAM & Home GPU Electricity vs. Cloud Calculator
Calculate how much GPU memory your LLM requires across FP16, 8-bit, and 4-bit quantization (including KV cache context overhead) and compare your local PC's power bill against on-demand Cloud GPUs.
Interactive LLM VRAM Calculator
Home GPU Electricity Bill Calculator
🖥️ Check Your Local GPU in Terminal (Zero-Install CLI)
Zero Dependencies · Python 3Auto-detects your NVIDIA GPU (nvidia-smi) or Apple Silicon Unified Memory and prints which Llama, Qwen, DeepSeek, Gemma, and Mistral models fit—with exact ollama run and vllm serve commands. Supports --vram 24 --ctx 32768 --live.
curl -sL https://stigla610.github.io/ai-cost-engine/tokenbench.py | python3 -
irm https://stigla610.github.io/ai-cost-engine/tokenbench.py | python -
🛒 Local LLM Hardware Buying Guide (Best GPUs & Unified Memory Rigs)
Compare local GPU & Apple Silicon tiers by usable VRAM, 24/7 electricity draw, and model capacity before buying hardware.
NVIDIA GeForce RTX 4060 Ti (16GB VRAM)
7B to 14B models (Llama 3.1 8B, Qwen3 14B, Phi-4, Gemma 3 12B) in 4-bit Q4_K_M
NVIDIA GeForce RTX 3090 / RTX 4090 (24GB VRAM)
14B to 32B models (Qwen3 32B, Gemma 4 31B, Mistral Small 24B, Codestral 22B)
NVIDIA GeForce RTX 5090 (32GB GDDR7 VRAM)
32B models at 32K–64K context & 8-bit Q8_0 24B models with 1.79 TB/s bandwidth
Apple Mac Mini M4 Pro (64GB) / Mac Studio (128GB Unified)
70B to 123B models (Llama 3.3 70B, Qwen2.5 72B, Mistral Large, Llama 4 Scout) at ultra-low 90W power
Open-Weight & Popular Model VRAM Guides
Select any model below to view quantization memory tiers and check if your GPU can run it.