Researchers from CEA-Leti, Spintec and C2N (Université Paris-Saclay) have developed a hybrid nanoelectronic Ising machine that combines hafnium-oxide RRAM with stochastic magnetic tunnel junctions (SMTJs). Such a machine searches for the lowest-energy state of a network of interacting binary variables, which makes it a good fit for hard combinatorial optimization problems such as vehicle routing, power-grid balancing and chip layout.
In the new system, an RRAM crossbar array encodes the graph of the problem, while the SMTJs, whose magnetization flips randomly under ordinary thermal noise, act as probabilistic bits. The two are connected in a single loop, so the system moves toward better solutions without a digital controller managing every update. The researchers also found that tuning the read voltage across the RRAM array gradually suppresses the randomness of the SMTJs as the search progresses, which provides the annealing behavior intrinsically, without extra circuitry.
Operating at room temperature and with no applied magnetic field, the prototype reliably found the global optimum of a 24-vertex weighted MAX-CUT problem and a 10-vertex, three-color graph-coloring problem. Updates are currently sequential, but the researchers say that nanosecond-scale device switching, and connecting the crossbar directly to the SMTJs (removing analog-to-digital conversion), could speed the system up by orders of magnitude.
Both device types are CMOS-compatible and can be integrated in the back-end-of-line, which opens the door to 3D stacking for larger and denser implementations. Next, the team plans to move the control electronics closer to the devices, test larger problems, and benchmark the speed and energy efficiency against conventional hardware. The research was published in Nature Communications.
