Real World Maths Busters

The Maths Behind Everyday Life

The story, the formula and a worked example — see exactly where this maths shows up in the world around you.

Back to all stories
Technology 3 min read

Why a 7-billion-parameter AI needs 14 GB of memory

Each parameter takes two bytes, so model size in memory is just a multiplication.

People ask whether their graphics card can run a given AI model. The answer is mostly memory: each parameter needs a couple of bytes, so a 7-billion-parameter model wants roughly 14 GB just to load.

Parameters times bytes

Multiply the number of parameters by the bytes each uses (2 for 16-bit precision) to estimate the memory needed.

memory=params×bytes/param\text{memory} = \text{params} \times \text{bytes/param}

VRAM to load a model

A 7B model

At 16-bit precision.

Memory to load it

  1. 1

    Multiply out

    7×109×2=14 GB7\times10^{9} \times 2 = 14\text{ GB}
  2. 2

    Add overhead

    Activations and context push it higher — so a 16 GB card is the practical minimum.

Answer

About 14 GB to load — which is why VRAM, not speed, is the gatekeeper.

The formula behind it

Model Memory

params×bytes\text{params} \times \text{bytes}

The takeaway

Model memory is parameters times bytes per parameter. It is why VRAM capacity decides which AI models a GPU can run.

Back to all stories