LLM Parameter Calculator
Estimate the GPU memory needed to run or fine-tune an LLM by parameter count.
GPU Compatibility
| GPU | VRAM | Fits? |
|---|
How to Estimate LLM Memory
The base formula is: Model Memory = Parameters × Bytes per Parameter. For inference, you also need KV-cache memory for the sequence length. For fine-tuning, multiply by 3-4x for gradients and optimizer states (8-12x for full fp32 Adam).
Important Notes
- Client-side processing only — no data sent to server.
- These are minimum estimates. Real deployments need extra memory for OS, CUDA kernels (~1-2GB), and frameworks.
- LoRA fine-tuning typically requires ~1.2x base model + adapter overhead.
- Free to use with no login/signup required.
- Report bugs in comments with sample input and expected output.
Comments & Discussion
Facing issues? Have questions? Post them here! We're happy to help!