Conventionally, resistive RAM memory is manufactured in a cross-point architecture. A cross-point at which a memory cell is located, is an intersection of vertical and lateral metal lines for the active and inert electrodes of the resistive memory cell. Here, we investigate the thermally induced degradation of Cu/TaOx/Pt/Ti ReRAM devices. When one cell is repeatedly switched on and off, a certain amount of heat is being deposited in the cell. The local temperature to rupture a Cu filament has been estimated to be 600-800oC. The lingering heat affects not only the device itself, but the heat dissipated along the electrode metal lines causes performance degradation of the neighboring cells. To monitor the cell degradation we choose specific set condition to form a marginal, i.e. weak, highly resistive Cu filament by imposing low Icc of 10μA and ramp rate rr=1.2V/s, while the reset is performed at low rr=0.1V/s with no Icc imposed in order to maximize the Joules heating. Such a fragile Cu filament acts as a canary in the coal mine with respect to ambient heat. The heat deposited in a stressed device leads to a limited number of switching cycles, usually 11-14. When the device is preheated, the maximum number of cycles may decrease to zero and thus serves as a measure for performance degradation of the neighboring cell. Within 3-4 minutes, after reaching maximal cycles for a given device, we test the switching behavior of the neighboring cells, one at a time. We find that only the cells that share either metal line of the heated device, suffer performance degradation. Direct neighbors (like a cell at the intersection of the adjacent Cu and Pt line) to the heated cell thus not sharing any the electrode of the heated device, are not affected at all by the heat, provided that the intermediate cells are in off-state. However, when the intermediate cells are programmed to the on-state and provide thus a direct heat conduction path to the heated device via copper line, Cu filament of the intermediate cell, and the Pt line, the diagonal neighbors suffers performance degradation. We find that the neighbors along the common Cu electrode line are affected more than the neighboring cells along the Pt electrode line. The heat transport along the Cu layer line is more effective and reaches further neighbors than for Pt line. Although the heat conductivity of Cu is roughly 5 higher than of Pt and the Cu metal line is 150 nm while the Pt line is 50 nm thick, we found this result surprising because it challenges the conventional assumption that the shape of the Cu filament is that of a sharply tipped cone with a broad base forming an interface with the Pt line and the tip of the cone touching the Cu line. This shape would imply very small contact of the filament with Cu and large contact with Pt, favoring heat transfer to the Pt line rather than to the Cu line. The observed more efficient heat transfer from the filament to Cu than to Pt line implies that the shape of the Cu filament appears to be more likely that of hour glass rather than a cone. When the tested neighboring cells are allowed to cool off for sufficiently long time (20 min or more) they return to the original performance levels of 11-14 cycles. We find also that the minimum cooling off period is shorter for cells disposed along the Cu lines rather than along the Pt lines.
The width of Cu and Pt lines in our arrays varies between 5 μm and 35 μm and the distance between the lines is ca 150 μm. We find more pronounced degradation of the neighboring cells along the Pt lines, for 35 μm wide Pt lines than for 5 μm wide Pt lines, indicating that metal lines with smaller cross-section are less effective in heat dissipation to the surroundings. In commercial arrays, the thickness, width, and line pitch of the electrode metal lines are of the order of couple 10s of nm. We therefore expect much more pronounced heating effects and over a longer time period in commercial memory arrays.
5:00 PM–7:00 PM Apr 23, 2019 (US - Arizona)
PCC North, 300 Level, Exhibit Hall C-E