AI hardware sizing
GPU VRAM Estimator
Estimate the memory footprint of model weights from parameter count, bits per parameter, and a configurable overhead factor.
0 GB
Weights-only estimate
Limit: this is a weights-memory estimate. Actual inference/training memory also depends on activations, KV cache, optimizer states, runtime buffers, batching, sequence length, sharding, and framework behavior.
Base equation
weight memory ≈ parameter_count × bits_per_parameter ÷ 8
For a 7-billion-parameter model at 16-bit precision, the raw weight requirement is roughly 14 GB in decimal units. The tool then applies the overhead factor you choose.
Why quantization matters
Reducing bits per parameter reduces the memory used by weights. That can make larger models fit on a given GPU, but lower precision can introduce quality or compatibility trade-offs, and it does not eliminate runtime memory requirements.
Read the full guide: How to estimate GPU VRAM for an AI model