Accessing VRAM memory with CPU

[LemonParty]

I have a technical question.
Today we have a lot of systems with unified memory.
Is it possible to access VRAM part of RAM with CPU directly?

[marcov]

You can map it into address space, this is commonly done for uploading textures. It might not be as fast (since it maybe has to go over the PCIe bus if it is an external card) though.

[LemonParty]

Understood. Thank you.

[JollyHong]

On unified memory architectures (like Apple Silicon or modern APUs), CPU and GPU share the same physical RAM, so there isn't really a physical transfer over a PCIe bus. However, the OS and graphics API still enforce memory domain/mapping restrictions.
If you are using APIs like DirectX 12, Vulkan, or Metal, you can create staging buffers/resources flagged with host-visible (or CPU-accessible) memory properties. This lets the CPU write or read directly using mapped memory pointers without requiring an explicit copy command, though cache coherence rules still apply depending on the architecture.
What graphics API or OS framework are you using for your project?

[LemonParty]

JollyHong, currently I am not developing projects with graphics API. But I was curious if CPU can combine computational powers with GPU.

[Thaddy]

But it CAN: FreePascal supports OpenCL.