Table 4.

Results summary from RSNN on Loihi and RSNN, eLSTM and LSTM on Jetson for accuracy, total power, energy per sample, delay and energy-delay product.

	Results summary				Comparison with RSNN on Loihi
Network	RSNN	RSNN	eLSTM	LSTM	RSNN	eLSTM	LSTM
Hardware	Loihi	Jetson	Jetson	Jetson	Jetson	Jetson	Jetson
Input	Events	Events	Events	Frames	Events	Events	Frames
Accuracy (%)	78.32	79.90	82.31	96.92	+1.58	+3.99	+18.60
Total power (mW)	31	3,851	7,642	5,385	124×	247×	174×
Total energy per sample (μJ)	71	1,108,695	96,000	35,212	15,615×	1,352×	496×
Delay per sample (ms)	2.3	295.3	12.9	6.7	172×	5.6×	2.9×
Energy-delay product (μJ·s)	0.16	327,398	1,238	236	2,046,237×	7,738×	1,475×

The number of trainable parameter (i.e., synaptic weights) are similar between the RSNN (236,700), the LSTM (225,975), and the eLSTM (236,919). Event-based inputs are encoded with threshold ϑ = 1. Comparisons with respect to RSNN on Loihi are evaluated as differences for the accuracy and as ratios for all the other quantities.