AI hardware · estimates, not benchmarks
Size the hardware, before you buy or download.
Small calculators for running language models on your own machine. Every result shows its formula, cites where it comes from and says it is an estimate.
LLM VRAM Calculator
GPU memory for a local model: weights, KV cache and overhead
KV Cache Size Calculator
Cache memory per token and for a full context, f16 to q4_0
Max LLM Context for Your VRAM
Tokens of KV cache that fit after the weights and a reserve
GGUF Quantization Size Comparison
Q2 to FP16 file sizes from a parameter count
LLM Tokens per Second Estimator
Decode speed ceiling from memory bandwidth and model size
Electricity Cost per Million Tokens
Local LLM running cost from watts, speed and kWh price
Local LLM vs API Break-Even
Months until hardware beats paying per million tokens
Multi-GPU VRAM Calculator
Pooled VRAM, fit check and tensor split for several cards
Model Download Time and Disk Planner
Download time and disk fit for a local LLM file
What these estimates are for
These tools turn model architecture numbers and your own hardware figures into planning estimates. They are not benchmarks: real memory use and speed depend on the model file, the runtime, drivers and the machine. Check your runtime's own output before you buy anything.