Table 3.

Programmibility (flexibility) of different dimensions in different neuromorphic processors.

Architecture	Mapping	Data-type	Network	Neuron	Synapse	Energy per
			architecture	model	model	SOp (pJ)
ODIN	Low	Fixed	Fixed	Fixed	Fixed	12.7
(Frenkel et al., 2018)
ReckOn	Low	Fixed	Fixed	Fixed	Fixed	5.3
(Frenkel and Indiveri, 2022)
μBrain	Low	Fixed	Fixed	Fixed	Fixed	26
(Stuijt et al., 2021)
TrueNorth	Low	Fixed	Low	Fixed	Fixed	2.5
(Akopyan et al., 2015)
Tianjic	Low	Fixed	High	Medium	Fixed	1.54
(Deng et al., 2020)
NeuronFlow	Low	Low	Medium	Medium	Fixed	20
(Moreira et al., 2020)
Loihi	Low	Low	Medium	Low	Medium	23.6
(Davies et al., 2018)
Loihi2	High	Medium	Medium	High	Medium	NA
(Davis, 2021)
SpiNNaker	High	Medium	High	High	High	45
(Stromatias et al., 2013)
SpiNNaker2	High	Medium	High	High	High	10
(Höppner et al., 2021)
SENECA	High	Medium	High	High	High	2.8

Synaptic Operation (SOp) varies in different applications and is only mentioned for high-level comparison. Mapping: low—hard partitioning of memory for weight and state; high—flexible memory reusing. Data-type: fixed—single data type supported; low—limited data type supported and only support binary events; medium—mixed-precision data type supported and graded events supported. Network architecture: fixed—only support Fully-Connected network; low—optimal support on Fully-Connected network and very costly support on CNN; medium—optimal support on Fully-Connected network and costly support to CNN; high—optimal support to both fully-Connected and CNN, and can also support novel network architectures. Neuron model: fixed—single fixed model; low—single predefined model with limited programmability; medium—multiple predefined models with limited programmability; high—fully programmable model. Synapse model: fixed—single fixed model; medium—single fixed model with limited programmable learning support; high—fully programmable model and fully programmable learning support.