Despite the large video memory capacity, its video memory bandwidth is very low. I guess the model's decode speed will be very slow. Of course, this design is very well suited for the inference needs of MoE models.