sure on consumer GPUs but that is not what is constraining the model inference in most actual industry setups. technically even then, you are CPU-GPU memory bandwidth bound more than just GPU memory, although that is maybe splitting hairs