Table 1. Performance difference across SC and conventional binary domain.

CNN Model	Platform	Year	Method	Area (mm2)	Power (W) or energy (nJ)	Accuracy (%)	Energy efficiency (images/J) or (GOPS/W)
Lenet-5	CPU	2009	Software	263	156 W	99.17	4.2
		GPU	2011	Software	520	202.5 W	99.17	3.2
		ASIC	2016	SC 256 bit (Ren et al., 2017)	36.4	3.53 W	98.26	221,287
		ASIC	2018	SC 128bit (Li et al., 2018a)	22.9	2.6 W	99.07	1,231,971
	ASIC	2018	SC DWM 128bit (Ma et al., 2018)	19.8	0.028W	98.94	–
AlexNet (last second layer)	CPU	2009	Software	263	156 W	–	0.9
		GPU	2011	Software	520	202.5 W	–	2.8
	ASIC	2018	SC 128bit (Li et al., 2018a)	24.7	1.9 W	–	1,326,400
Custom (3x3filter)	ASIC	2015	Binary	5.429	3.287mW	–	–
		ASIC	2017	SC MAC	1.408	1.369mW	–	–
	ASIC	2019	SC DMAC	1.439	1.393mW	–	–
Custom (Ardakani et al., 2017)	ASIC	2017	Binary	–	380 nJ	97.7	–
	ASIC	2017	Integral SC	–	299 nJ	97.73	–
ConvNet for MNIST	ASIC	2015	Binary	0.98	0.236W	–	1158.11 GOPS/W
	ASIC	2017	SC MAC	0.43	0.279W	–	5640.23 GOPS/W
Custom (Hirtzlin et al., 2019)	ASIC	2019	BNN	1.95	220 nJ	91	–
	ASIC	2019	SC BNN	0.73	90 nJ	89.6	–

Notes.

GOPS

Giga operations per second