Pipelined Training with Stale Weights in Deep Convolutional Neural Networks

Table 6

Inference accuracy and speedup of actual pipelined/hybrid training for CIFAR-100.

ResNet	PPV	Accuracy			Time (seconds)			Speedup
		Nonpipelined (%)	Pipelined (%)	Hybrid (%)	Nonpipelined	Pipelined	Hybrid	Pipelined	Hybrid
-56	(19)	72.30	70.25	72.34	10,068	6,040	8,113	1.67X	1.24X
-110	(37)	72.63	71.52	73.03	19,624	11,162	15,591	1.76X	1.26X
-224	(75)	73.00	70.90	72.69	40,807	22,561	32,132	1.81X	1.28X
-362	(121)	71.77	71.03	72.09	66,977	37,008	51,688	1.81X	1.29X