ReRAM device fabrication and test

We fabricated 2-terminal oxide ReRAM with device dimensions of 50 × 50 μm². First, a SiO₂ underlayer was grown on a 200 mm Si wafer. Then, a 100 nm-thick TiN film was deposited by reactive sputtering as a bottom electrode, followed by deposition of a HfO₂ layer by atomic layer deposition as a switching layer where a current conducting filament is formed. We varied the thickness of the switching layer (device A: 5 nm, device B: 4 nm) to investigate its impact on switching symmetry. Next, a 20 nm-thick TiN was deposited by reactive sputtering as a top electrode. The device area was defined by photolithography and reactive ion etching of the TiN electrode. To test switching symmetry and SNR of our ReRAM devices, we applied a sequence of weight update (write) pulses with the same voltage amplitude for each polarity. We used high-resolution source measure unit (SMU) to read the device conductance state between the write pulses. We applied a small read voltage of 0.1 V to prevent disturbance in the resistance state. While keeping the read voltage applied across the device, we took multiple read steps with a 16.67 ms integration time until the measured values read at the instrument stabilized (typically within 3–10 repetitive read measurements in the device resistance range of interest). Then, we chose the last measurement as the representative value. We did not detect random telegraph noise with this read sequence. The write pulses had duration of 100 ns (unless otherwise mentioned) and various voltage amplitudes (set pulse: 1.6–1.7 V; reset pulse: −1.8 to −1.9 V) were compared to investigate the impacts on switching symmetry and SNR. In order to separate noises from weight update and those from weight read, we also carried out read-only test, where only read steps were repeated up to 1000 times without weight updates in between. Our linear regression analysis showed that the residual standard error of read-only trace is 2.51 × 10⁻⁷ S, which is almost one order lower than that of read-after-write trace (1.38–1.57 × 10⁻⁶ S). Therefore, we attribute a majority of noise components of our ReRAM devices to inherent randomness in weight updates.