It has a unified memory architecture, so the 128G is shared directly between CPU and GPU. Though it's slower than dGPU VRAM.