DDR5 memory has higher bandwidth generally 5x to 10x that of DDR3, while DDR3 memory has higher capacity (10s GB vs TB).

A discrete GPU is separate from the CPU. I'm curious how this changes communication compared to other types of GPUs. Would CUDA still work on these?

CUDA should still work, I don't see why not. I think this is just how most systems are set up though nowadays. The discrete GPU is a separate processing unit that is installed in one of the PCIe slots on a motherboard as shown. It's just a dedicated/discrete graphics card for graphics compute-intensive applications as is common in most laptops/PCs in addition to an integrated graphics that is used to run low-intensity graphics applications.

It seems like CUDA is an nvidia specific programming abstraction, so CUDA doesn't even work on AMD GPUs. So I doubt CUDA is going to be available for integrated graphics or any other non nvidia GPU.

I would imagine that it could in theory be possible to write a compiler that can run code for e.g. AMD GPUs, but I do not think anyone will do so, given how there exists openCL to essentially serve as a general computing extension to C and C++

Look up NVIDIA Optimus for an example of a userland software tool that assists with discrete GPU switching on laptops. Optimus allows the user to dynamically switch between discrete GPUs and integrated graphics depending on when energy needs are important.