Table 2.

Comparison of MNIST benchmark results across neuromorphic platforms

	Platform	Reference	Architecture	Node, nm	Accuracy, %	Energy/inference, µJ	Throughput, inference×s–1	Latency, µs
Digital	SpiNNaker	Stromatias et al. (61)	784-500-500-10	130	95.0	/*	/*	/
	TrueNorth	Esser et al. (62)	CNN (1 ensemble)	28	92.7	0.27	1,000	/
			CNN (16 ensembles)	28	95	4	1,000	/
			CNN (64 ensembles)	28	99.4	108.0	1,000	/
	—	Chen et al. (63)	236-20	10	88.0	1.0	6,250	/
			784-1024-512-10	10	98.2	12.4	/	/
			784-1024-512-10	10	97.9	1.7	/	/
	MorphIC	Frenkel et al. (64)	784-500-10†	65	97.8	205	/	/
			784-500-10†	65	95.9	21.8	250	/
	SPOON	Frenkel et al. (38)	CNN	28	97.5	0.3‡	/	117
Analog	BSS-1	Schmitt et al. (56)	100-15-15-5	180	95.0	/*	10,000	/
	BSS-2	Göltz et al. (57)	256-246-10	65	96.9	8.4	21,000	¡10
	BSS-2	This work	256-246-10	65	97.6	2.4	85,000	8

^*Estimates were given by Pfeiffer and Pfeil (40).

^†Segmented input and hidden layers.

^‡Based on presilicon estimates.