Getrennte Adressräume

The CPU and GPU each own their own RAM. They do not share one address space, and that single fact shapes all of CUDA. 🧠

A pointer is just a number naming a slot in memory. That slot only makes sense in the address space where it was created.

A pointer from malloc names host memory. The GPU has no such slot, so handing it to a kernel points nowhere valid.

A pointer from cudaMalloc names device memory. The CPU cannot read it directly; dereferencing it on the host crashes.

Passing a host pointer into a kernel compiles fine but fails at runtime. The number is valid; the memory it names is not on the GPU.

Both pointers look like the same type to C++, so the compiler stays silent. The mismatch only bites when the GPU touches the address.

To move data between spaces you must copy it explicitly with cudaMemcpy. There is no automatic sharing across the divide.

Many programmers prefix host pointers with h_ and device pointers with d_. It is a habit that prevents painful mix-ups.

Inside a kernel every pointer you dereference must live in device memory. That is the only space the GPU threads can reach.

Later you will meet unified memory, one pointer valid on both sides. For now, keep host and device pointers strictly apart.

Picture two rooms with no shared shelves. To use data in the other room you must carry a copy over, never just point across.

Let us confirm you understand separate address spaces.

You learned host and device live in separate address spaces, pointers are not interchangeable, and cudaMemcpy is how data crosses the gap. ✅