Resistive RAM memory are manufactured in a cross-point architecture. A cross-point, where a memory cell is located, is an intersection of vertical and lateral metal lines for the active and inert electrodes of a resistive memory cell. The two metal electrodes are separated by a dielectric such as TaO_x. We have manufactured resistive ReRAM cells Cu/TaO_x/Pt/Ti and Cu/TaO_x/Rh/Cr. The thickness of TaO_x is 25 nm. The thickness of the Cu, Pt, Ti, Rh, and Cr layers is 150 nm, 50 nm, 20 nm, 50 nm, 20 nm, respectively. All layers have been deposited using e-beam PVD. The line width of the Cu, Pt/Ti and Rh/Cr varies from 5 μm to 35 μm. The distance between the lines is ca 150 μm. The latter parameter is a characteristic distance for heat transfer between the cells. Fresh isolated Pt/Ti and Rh/Cr devices display similar switching properties, with very similar V_form, V_set, V_reset, and R_on distributions. The inert electrode of Rh/Cr, has roughly twice larger heat conductivity than Pt/Ti. The heat conductivities for Pt, Ti, Rh, and Cr are 72, 22, 150, and 94 W/(mK), respectively, much inferior to the heat conductivity of Cu at 385 W/(mK).

When one cell in the array is repeatedly switched on and off, a certain amount of heat is being deposited in the device. The switching heat is dissipated to the environment mainly by the electrode metal lines. Our results show that the heat deposited in a device lingers for some time in the vicinity of the device and spreads to neighboring devices via the common electrode lines. Experiments show that just after repeated set-reset cycles of one device, the performance of neighboring devices is degraded when they have a metal line in common with the heated device. The question became how, in absence of a temperature probe, to measure such heat transfer and the resulting device degradation quantitatively. To probe into such degradation we have chosen specific set conditions of the device at a compliance current (I_cc) of 10 uA. For lower I_cc, the on-state is volatile, and for larger I_cc the cell is non-volatile and the filament is stable, Thus, I_cc=10 uA is a stability border case. A fresh device under such condition can be switched in sequence maximally 13-15 times. We find that for devices affected immediately after heating of a neighboring device this maximal number of switching cycles decreases significantly. We have measured the maximum number of switching cycles for neighboring devices along the Cu and along the Pt electrode lines. Because of the higher heat conductivity, the performance degradation for a Cu line extends to more distant neighboring cells than in the case of Pt/Ti electrode lines. After some cooling off period the devices displaying performance degradation recover quickly to the pre-stress performance level. We find that the cooling off period is shorter for the Cu lines than for the Pt/Ti lines. We observe also that the performance impact on the neighboring cells is larger for lines with small cross section (thickness × width) than for those with large cross section. However, for a short time (couple of minutes), the fast heat dissipation extends over a larger distance within which the devices may be affected by the heated device. This poses a dilemma in terms of the sequence, allowable frequency and location of the cells to be programmed. Interestingly, cells which do not share neither of the two metal lines with the heated device (even if geometrically close to it), are mostly unaffected by the heating of the stressed cell, provided there is no direct heat conductor path connecting the device with heat source. This condition is fulfilled when the intermediate cells lying in the path between the monitored device and the heated device are in the off-state. However, if the intermediate cells are in the on-state, they provide a direct electrical (thus a direct heat) path from the heated device to the tested device, leading to the performance degradation of the probed cell. The heat conduction is also affected by the cross-sections of the metal lines.
In commercial application the width and thickness of such metal lines are on the order of couple 10's of nm rendering the commercial arrays 100 times tighter in lateral dimensions than our samples. In nanowires the heat transport is no longer diffusive but ballistic which is known to make for a less efficient heat transport. Hence, for commercial memory arrays the nonlocal heating effects and the resulting device performance degradation are likely be more pronounced.