Skip to content
BENIBIX Learn

Unified Memory vs Video Memory (VRAM)

People think unified memory is just video memory with a new name, and it is not. Video memory sits on a graphics card for the GPU alone. Unified memory is one pool that the CPU and GPU share. So a 64 GB pool is not 64 GB for the GPU, since the system and apps use it too.

Infrastructure🔧 Hardware

Unified Memory

One pool of memory that the CPU and the GPU both use. Nothing is copied.

Unified memory is one pool of memory that the CPU and the GPU both use. In most computers, the GPU has its own video memory. Data is copied across to it. With unified memory there is no copy, because both chips read the same bytes. The memory sits right beside the chip, so it usually cannot be added to later. A big pool lets a home computer load AI models too large for most graphics cards.

Example
Emma's Mac has 64 GB of unified memory. She loads an AI model that needs 40 GB, which her old 24 GB graphics card could not hold.
Why it matters
Copying data to the GPU takes time and uses the space twice. A shared pool saves both.
Full entry
Infrastructure🔧 Hardware

Video Memory (VRAM)

The GPU's own fast memory. A model must fit in it to run on the GPU.

Video memory is the fast memory on a graphics card that only the GPU uses. It sits right next to the GPU chip. For the GPU, it is much faster than normal RAM. To run an AI model on a GPU, the model's weights usually have to fit in its video memory. Home graphics cards often have 8 to 32 gigabytes, and data centre GPUs have 80 or more.

Example
Emma's graphics card has 12 GB of video memory. A model with 7 billion weights at 2 bytes each needs about 14 GB, so it does not fit.
Why it matters
To run AI at home, video memory is most often the limit. It decides the biggest model you can load.
Full entry

In one line

Unified Memory
One pool of memory that the CPU and the GPU both use. Nothing is copied.
Video Memory (VRAM)
The GPU's own fast memory. A model must fit in it to run on the GPU.