GPU VRAM Calculator
How much GPU memory a model needs for inference, and which GPUs can hold it. Pick a Hugging Face model or enter a parameter count.
or
Memory needed
| Precision | Weights | Needed to run |
|---|
Which GPUs fit it
| GPU | VRAM | Fit |
|---|
How this is calculated. Weights = parameters × bytes per parameter (FP32 4, FP16/BF16 2, FP8 1, INT4 0.5). Needed to run = weights × 1.2, a common rule of thumb for KV cache, activations and allocator overhead at single-request inference around an 8K context. Large batches or long contexts can need much more, because the KV cache can exceed the weights. Figures are in GiB (1024³ bytes). Multi-GPU counts assume the model is split evenly with tensor parallelism.
Same formula as the VRAM calculator on aquanode.io, which also shows live GPU prices.