So how does training performance work when you don't have enough memory on the GPU? If you just use RAM instead, is the performance hit as simple as PCIE_Bandwi

So how does training performance work when you don't have enough memory on the GPU? If you just use RAM instead, is the performance hit as simple as PCIE_Bandwidth divided by GPU_Memory_Bandwidth? Example with 4090 would be 64GB/s and 1008GB/s respectively, so one could expect about 6% performance even when dealing with a model that is 48GB large (or more)? Or are there second-order effects that cut this speed down even more?