The portion of the DAG above the 4GB vram limit is stored in the system/pc ram, and DAG requests to the above-4GB limit are coursed through the PCIe data lanes since CUDA allows for consolidated access to system/pc ram via a feature called "Unified Memory". This is why the hashrate drops since all accesses to DAG data above the 4GB limit need to go through the PCIe lanes to the system/pc ram, and this is also why using a x16 PCIe slot improves the gpu's access to system ram.