Table 1 Comparison of state-of-art Ising computers

	CPU	Quantum annealers	CIM	Mem-HNN	PTNO	SMTJ-based Ising computer
Spin form	Artificial Ising spin	Qubits	Optical parametric oscillator	Memristor	Nano-oscillators	MTJ
Algorithm	Simulated annealing	Quantum annealing	Coherent computing	Modulate intrinsic noise annealing	Simulated annealing	Intrinsic noise annealing + global annealing
Connectivity	All-to-all	Sparse	All-to-all	All-to-all	All-to-all	All-to-all
Room temperature	Yes	No	Yes	Yes	Yes	Yes
Number of nodes solved experimentally (simulated)	4761	61	2000	60	8 (100)	4761
Number of devices (bits)	~10⁹	2048	1 (2048)^d	3600	8	80	4096
Power	65 W	25 kW		120 mW	2.56 mW^h	Experiment	Simulation using SMTJ in this work	Simulation using SMTJ in ref 45^m
Power	65 W	25 kW		120 mW	2.56 mW^h	0.64 mW^j	0.288 mW^l	0.288 mW
Size of the computer (chip)	14 cm²		3\(\times {10}^{4}\) cm³	64 cm²		8 cm²	12288 µm²	12288 µm²
Main frequency	3 GHz^a		200 kHz^e	1 MHz^g	301 Hzⁱ	10 kHz	10 kHz	1.1 MHz
Time to solution (TSP70)	12 s^b	>10⁴ s^c	80 ms^f	320 ms	1.3 × 10³s	40 s^k	50.54 s	0.64 s
Energy to solution (TSP70)	780 J	>2.5\(\times\)10⁸J		3.84\(\times\)10⁻²J	3.4 J	2.56\(\times\)10⁻² J	1.456\(\times\)10⁻² J	1.8\(4\times\)10⁻⁴ J
Energy efficiency (1/energy to solution)	1.3\(\times\)10⁻³	<4.0\(\times\)10⁻⁹		<2.6\(\times\)10¹	2.9\(\times\)10⁻¹	3.9\(\times\)10¹	6.8\(\times\)10¹	5.4\(\times\)10³

Unavailable quantities are left blank.
^aEstimated average main frequency of Intel Core-i7 PC.
^bSolution time using dataset st70 from ref. ⁷ by using TSP-SAGEP algorithm.
^cApproximation from the result of N = 55 node MaxCut problem.
^d1-km fiber with 2048 pulses per second used as Ising spins in a time-multiplexed way.
^eRound-trip time of 5 µs in experiment.
^fLinear approximation from the simulated result of N = 2000 node complete graph problem. The time of transferring data from FPGA to computer is 60 s×n, where n is the observation time.
^gMain frequency of on-chip computing array using 180-nm CMOS-memristor chip in experiments.
^hEstimated power consumption of main compute kernel.
ⁱEstimated for 80 node system.
^jExperimentally demonstrated power consumption of main compute kernel.
^kExperimentally tested total solution time using dataset st70 from TSPLIB including computing time, data transfer time, and updating time.
^lPower consumption of main compute kernel of 4 Kb SMTJ Ising computer (4-bit bit line) based on the experimentally calibrated SMTJ model.
^mSimulation results using SMTJ with a retention time of 8 ns in ref. ⁴⁵.

Search