VRAM Calculator for AI Models

Calculate exactly how much GPU memory you need to run any LLM or AI model.

Model Size (Billion Parameters)

Quantization (Precision)

FP16 / BF16

16-bit · 2 bytes/param

INT8

8-bit · 1 byte/param

INT4

4-bit · 0.5 bytes/param

Context Length (Tokens)

Bias Factor (GPU utilization efficiency)

19.2

GB VRAM Required

Llama-3.1-8B at FP16 with 4K context

GPUs That Can Run This

How this works: Model weights use params × bytes_per_param. KV cache adds 2 × layers × hidden_dim × context × bytes_per_param. We estimate hidden dimensions from parameter count using Chinchilla scaling. Bias factor accounts for fragmentation, CUDA kernels, and runtime overhead. Real-world usage may vary ±15%.

Quick reference: 7B FP16 ≈ 14GB · 7B INT4 ≈ 4GB · 13B FP16 ≈ 26GB · 13B INT8 ≈ 13GB · 70B FP16 ≈ 140GB · 70B INT4 ≈ 35GB · 405B FP16 ≈ 810GB