Which GPU do you need? Find out in 10 seconds.

Pick what you want to do and how big your model is. We work out the GPU memory you need, which NVIDIA GPU fits, how many, and what it costs per month.

1. What do you want to do?

2. How big is the model?

Or type the size in billions of parameters 70B

3. A few details

Precision (for running the model)

Plan

Hours per month 730 h

How we work it out

Running a model

Each parameter takes 2 bytes at 16-bit, 1 byte at 8-bit and half a byte at 4-bit. We add about 20% for the attention cache and runtime.

LoRA fine-tuning

The base model stays in 16-bit and only small adapter layers are trained, so you need roughly the model size plus about a third.

Full fine-tuning and pre-training

Training with the Adam optimiser in mixed precision needs about 16 bytes per parameter, plus room for activations.

This is a planning estimate. Long context windows, large batches and your framework change the real numbers. Our engineers check your setup before you buy. Read the full GPU memory guide →