1. Artificial Neuron Based on TiO2 Cbram for Neuromorphic Computing
- Author
-
Dong-Won Kim, Ki-Hyun Kwon, Hea-Jee Kim, Soo-Min Jin, Hun-Mo Yang, Ji-Yeon Kim, and Jea-Gun Park
- Abstract
Recently, CMOS-based artificial neurons have successfully implemented neuron functions such as integrate and fire (IF) or leaky integrate and fire (LIF). In general, CMOS-based artificial neurons use capacitors to implement integration function of neuron. In this case, required RC time constant is about few milliseconds. And it requires large area of capacitor more than 1000F2. Therefore, it is necessary to develop capacitor-less artificial neuron to achieve high neuronal density. In this works, we implemented the integration function of neurons using CBRAM instead of capacitors. For the neuron device, CuTe/TiO2/TiN CBRAM cell was fabricated as shown in figure 1a. The neuron device has W bottom electrode with ~113 nm2 pattern size and 10 nm-thick TiO2 layer, and 200 nm-thick CuTe top electrode with 60 μm2 size. The TiO2-based CBRAM showed a typical bi-stable I-V curve as shown in figure 1b. The CBRAM cells showed negative differential resistance (NDR) during reset process. And it showed integrate characteristic when consecutive negative voltage pulse was induced to top electrode of the CBRAM cell. Figure 1c shows integrate characteristic of TiO2-based CBRAM cell. As shown in figure 1c, as the voltage pulse is applied, current gradually decreases. Also as amplitude of voltage pulse increased, the change of current became greater. Because the area of the CBRAM cell is only 4F2, it is expected that area of artificial neuron can be significantly reduced by replacing conventional capacitor based neuron. We will present how the CBRAM cell can be used to implement artificial neuron. Figure 1
- Published
- 2019
- Full Text
- View/download PDF