Table 2.

Performance metrics for four different network configurations, and reductions in memory read volume realized by pipelined backpropagation.

Network configuration	Write volume (Gwords)	Read volume (Gwords)	Percentage reduction in read volume due to pipelined backpropagation	Average training time per example (ms)
16-bit weights. 0/1 hidden activations	3.24	314	15%	15.0
8-bit weights. 0/1 hidden activations	1.63	135	12%	12.0
16-bit weights. -1/1 hidden activations	6.03	663	36%	32.8
8-bit weights. -1/1 hidden activations	4.90	337	36%	32.5