Table 2.

Performance (GOP/s), power consumption (W), and energy efficiency (GOP/J) for different technologies candidated to accelerate the DeepSurv workload.

Technology	Data format	GOP/s	Power consumption	GOP/J
This work (IMC)	17 levels	1.82	17.1 mW	106
XC7Z045 (Qiu et al., 2016)	INT16	136.97	6.63 W	14.22
XC7Z020 (Venieris and Bouganis, 2016)	INT16	12.73	1.75 W	7.27
ZCU102 (Lu et al., 2017)	INT16	2940.7	23.6 W	124.6
XCVU440 (Shen et al., 2018)	INT16	785	26 W	30.2

The data shown for the field programmable gate array (FPGA) were obtained (Qiu et al., 2016; Venieris and Bouganis, 2016; Lu et al., 2017; Shen et al., 2018) and represent the maximum theoretically attainable values when a generic neural network task is run.