G$) assuming I$_{prog}$ $\sim$ 100 $\upmu$A, $G \sim$ 1 $\upmu$S, and $t_{pw} \sim$ 100ns, as per the parameters mentioned in⁵⁰. &Learning energy reported at V= 0.5V⁵⁹. $^*$Only coupling weights are PCM-based; eligibility computations are performed using a high-precision unit. Write energy is estimated roughly according to the values provided in⁴⁹: I$_{prog}$ = 700$\,\upmu$A, $t_{pw}$ = 600ns, $G = 10\, \upmu$S. $^\#$Learning energy dominated by optical power⁶⁰. $^\otimes$ Limited by PCM device resistance drift rate. $^\boxtimes$Limited by Von-neumann style sequential computing. $^\triangle$Limited by Ag conductive filament relaxation dynamics. $^\Omega$Limited by thermal time constant.

	Coupling weights	Eligibility	Per synapse area	Energy per timestep	Eligibility decay time constant	Maturity
Y. Demirağ et al.⁵⁰	PCM conductance	PCM drift	$12 \times 12 \, \upmu \textrm{m}^2$	$\sim 1 \, \textrm{nJ}^{!}$	$\sim \textrm{s}^{\otimes }$	+
C. Frenkel et al.⁵⁹	CMOS	CMOS	$1200F^{2 \, \$}$ (0.94 μm²)	$1.5-178 \, \textrm{nJ}^{ \& }$	$\sim \textrm{ms}^{\boxtimes }$	++
T. Bohnstingl et al.⁴⁹	PCM conductance	CMOS	$5221F^{2 \, \Lambda }$ (1.02 μm²)	$29 \, \textrm{nJ}^{*}$	$\sim \textrm{ms}^{\boxtimes }$	+
S. G. Sarwat et al.⁶⁰	PCM conductance	Optical response of PCM	$5 \times 4.8 \, \upmu \textrm{m}^2 \, \Upsilon$	$\sim \textrm{mJ}^{\#}$	$\sim \mathrm {ns{-}ms}^{\triangle }$	-
This work	ReRAM conductance	ReRAM local temperature	$450F^2$ (1.9 μm²)	$5 \, \textrm{pJ}$	$\sim \mathrm {ns{-}\upmu s}^{\Omega }$	-

Search