Models can split loads across a discrete GPU and CPU/RAM.

Its not as fast as if you can load it all in the GPU, but it gives you more options. Its been quite common for a long time.