What is the distinction between shared and dedicated GPU memory on a unified memory system?

[Sgraffite]

In this scenario I am using a mini PC with unified memory. In the BIOS I choose how much RAM to allocate to the GPU. There is no discrete GPU in the system.

Right now I have 96GB of RAM and 48GB is dedicated to the GPU according to the BIOS.
Windows shows there is 48.5GB of dedicated and 23.8GB of shared.

et

 

[Quintessa]

There is no discrete GPU in the system.

Exactly, so all GPU memory is still system RAM. The division you see is purely logical, driven by BIOS and OS memory managers. If you want more RAM available to apps, try dropping that 48GB reservation lower (e.g., 8 to 16GB) unless you're doing heavy GPU workloads. Most iGPU tasks (even 4K playback, some light 3D) won’t saturate even 4-8GB.

Right now I have 96GB of RAM and 48GB is dedicated to the GPU according to the BIOS.

When BIOS allocates 48GB as "dedicated", that portion is statically reserved for GPU use and unavailable to the OS as system RAM, even if the GPU isn't using all of it.

Windows shows there is 48.5GB of dedicated and 23.8GB of shared.

Here’s the distinction:
- Dedicated GPU memory: RAM physically reserved exclusively for the GPU (as set in BIOS).
- Shared GPU memory: RAM that remains accessible to both the CPU and GPU dynamically, used as overflow if the dedicated pool is full.

 

[Sgraffite]

Exactly, so all GPU memory is still system RAM. The division you see is purely logical, driven by BIOS and OS memory managers. If you want more RAM available to apps, try dropping that 48GB reservation lower (e.g., 8 to 16GB) unless you're doing heavy GPU workloads. Most iGPU tasks (even 4K playback, some light 3D) won’t saturate even 4-8GB.

When BIOS allocates 48GB as "dedicated", that portion is statically reserved for GPU use and unavailable to the OS as system RAM, even if the GPU isn't using all of it.

Here’s the distinction:
- Dedicated GPU memory: RAM physically reserved exclusively for the GPU (as set in BIOS).
- Shared GPU memory: RAM that remains accessible to both the CPU and GPU dynamically, used as overflow if the dedicated pool is full.

So you are saying that it could use the dedicated + shared (~72.3GB) amount for the GPU, even though the BIOS specifies 48GB?

I thought it would never go over the 48GB if that is what I set it to.

 

[Quintessa]

So you are saying that it could use the dedicated + shared (~72.3GB) amount for the GPU, even though the BIOS specifies 48GB?

Exactly. The 48GB BIOS setting is just the guaranteed reserved pool (i.e. "dedicated"). That RAM is locked for GPU use only, OS and apps won’t touch it.

I thought it would never go over the 48GB if that is what I set it to.

It can. If a workload demands more VRAM (say, large textures or GPU compute), the system can tap into the shared pool, i.e. regular system RAM dynamically up to the limits Windows and the hardware allow.
So real-world usable GPU memory = dedicated (BIOS) + shared (dynamic), which in your case is ~72.3GB total.

Just note: performance past the dedicated amount drops since access is no longer through high-speed, pre-allocated channels.