MRAM based neuromorphic cell for Artificial Intelligence
Published : 18 October 2019
Applications in artificial intelligence architectures require cell elements that can take multiple states based on different cell inputs. Hardware realizations of such elements can also be realized on magnetic non-volatile memory (MRAM) cells. This technology being developed at Spintec, associates non-volatility with fast switching of the order of a few nanoseconds. In conventional binary memory applications only two possible states are available for each cell. However, mimicking neuron behavior requires the possibility of multiple output states. This can be obtained based on individual MRAM pillar elements that are then connected as parallel or series resistors. The global measured resistance state will depend on the actual states of individual cells. The possibility to switch individual elements, while others also subjected to the same voltage maintain their original state relies on the natural dispersion of switching voltages. The goal of the internship will be to determine possible operation window to selectively switch individual memory elements and achieve stable multi-value outputs in the global cell. The sensitivity of parameters to control multilevel states will be investigated, focusing on pulse voltage and duration, applied magnetic field, and constant applied DC current. The temperature dependence of the cell characteristics will also be studied. Potential applications of this concept would be for example in memory computing that can be simulated once the electrical of multilevel cells is established.