User replaceable RAM is slow, which means you can’t integrate the CPU and GPU in one package. This means a GPU with it’s own RAM, which has huge disadvantages.
Even a 4090 only has 24GB and slow transfers to/from VRAM. The GPU can only operate on data in VRAM, so anything you need it to work on you need to copy over the relatively slow PCIe bus to the GPU. Then once it’s done you need to copy the results back over the PCIe bus to system RAM for the CPU to be able to access it. This considerably slows down GPGPU tasks.
User replaceable RAM is slow, which means you can’t integrate the CPU and GPU in one package. This means a GPU with it’s own RAM, which has huge disadvantages.
Even a 4090 only has 24GB and slow transfers to/from VRAM. The GPU can only operate on data in VRAM, so anything you need it to work on you need to copy over the relatively slow PCIe bus to the GPU. Then once it’s done you need to copy the results back over the PCIe bus to system RAM for the CPU to be able to access it. This considerably slows down GPGPU tasks.