38 GPU cores. The neural network inference cores don't do floating point and therefore aren't used for training or games. Assuming these GPU cores are comparabl

38 GPU cores. The neural network inference cores don't do floating point and therefore aren't used for training or games. Assuming these GPU cores are comparable, if >3x isn't a reasonable target, what would you suggest is?