Fig. 5: Scaling with problem size and cross-bar implementation towards large-scale Ising computing.

a Solution quality of various problems using different number of SMTJs (M) in the array. The datasets used are St70, Eil101 and KroA200, for 70, 101 and 200 cities, respectively. b Total length of KroA200 TSP at different convergence speeds using different number of SMTJs. The dashed line represents the best demonstrated solution. c Success probability of different TSP algorithm (without/with GP and CTSP) as the number of cities increases after running for 50 times. A total of 512 SMTJs are used. Here we define the success as achieving the solution quality of 95%. d SMTJ cross-bar array which contains row decoder, SMTJ, select transistor and read sense amplifier (RSA). BL represents bit line, WL represents word line, Vin, Vout and Vdd represent the input voltage, output voltage and supply voltage of RSA. e Circuit of one RSA which contains a current mirror, voltage equalization circuit (VEC, with a control signal of EQ which initializes the voltages in Q and QB points, under a reference voltage of Vdd/2), voltage sense amplifier (VSA, with a control signal of SEN), reference resistance (\({{{{{\rm{Rref}}}}}}=\frac{1}{2}({{{{{\rm{Rap}}}}}}+{{{{{\rm{Rp}}}}}})\), Rap and Rp represent SMTJ’s resistance in AP and P state respectively), and control transistors. f Signals of writing/reading two adjacent SMTJ cells in one BL, selected by WL0 and WL1 in sequence. All signals are defined in e and f.