Pipelined Training with Stale Weights in Deep Convolutional Neural Networks

Table 5

Inference accuracy and speedup of actual pipelined/hybrid training for CIFAR-10.

ResNet	PPV	Accuracy			Time (seconds)			Speedup
		Nonpipelined (%)	Pipelined (%)	Hybrid (%)	Nonpipelined	Pipelined	Hybrid	Pipelined	Hybrid
-56	(19)	92.63	92.89	92.75	6,745	4,090	5,429	1.65X	1.24X
-110	(37)	93.59	92.88	93.55	13,150	7,570	10,452	1.73X	1.26X
-224	(75)	92.77	91.39	93.33	27,231	14,998	21,245	1.81X	1.28X
-362	(121)	93.46	90.53	93.98	44,814	24,640	34,814	1.82X	1.29X