Increasing raw computing power is the primary factor in accelerating neural network training performance