Scalable methods for 8-bit training of neural networks