Supplementary information for A fast and low power microelectromechanical system based nonvolatile memory device Sang Wook Lee, Seung Joo Park, Eleanor E. B. Campbell & Yung Woo Park The supplementary information includes: 1. Supplementary Figures S1-S5 2. Supplementary Discussion: Description of the comparison of the energy consumption of our device with conventional flash memories 3. Supplementary reference
Supplementary Figure S1. Fabrication process of the memory device The CNT growth was performed using thermal CVD with 1000 sccm CH 4 / 100 sccm H 2 at 900 o C for 15 min. (a) Preparation of source, drain and actuating electrodes. The pattern was made using EBL. 0.5 nm Ti, 39.5 nm Pd and 50 nm Al 2 O 3 was deposited by e-gun evaporator. (b) Top-floating gate fabrication. (c) Side-floating gate fabrication. (d) Deposition of the suspended cantilever. Dose-variation exposure and critical point drying were used.
Supplementary Figure S2. Dose variation exposure for suspended structures (a) 4 layers of e- beam resists (PMMA and EL10) were coated on the Si wafer. (b)the cantilever structure was patterned with varying dose conditions of 170 ~ 450 μc/cm 2 using a JEOL JBX-9300FS EBL machine. (c) Pattern was developed by MIBK:IPA = 1:2 for 20 min. (d) Metallization process. 1 nm Cr / 200 nm Al / 0.8 nm Cr was deposited using e-gun evaporator. (e) Lifted-off in acetone and dried by critical point dryer. (f) FE- SEM images of point-contact structure.
Supplementary Figure S3. Transfer characteristics of CNT-FET. Backgate dependence of I ds for the side-floating gated CNT device (a) and the top-floating gated CNT device (b) Back gate on the Si substrate was swept from -10 V to 10 V with V ds = 100 mv in air at room temperature. Black arrows indicate sweep directions.
Supplementary Figure S4. Detection of the mechanical resonance frequency (a) The schematic of the measurement setup (optical interferometer method). (b) Resonance frequency measurement yielding a value of 8.5 MHz with a Q-factor of around 50.
Supplementary Figure S5. Detection of the switching speed of MEM cantilever (a) The schematic of the measurement setup. A 3 terminal micromechanical system which has direct connection between floating gate and an oscilloscope channel was prepared for this measurement. The fabrication procedure and dimensions of the cantilever were identical to the floating-gate memory devices. Switching time of 164 ns (b), release time of 181 ns (c), and circuit delay time of ca. 35 ns (d) 34 ns for the circuit on and (e) 36 ns for the circuit off) was detected in this measurement, giving actual switching time of 130 ns (on) and 145 ns (off).
Supplementary Discussion Comparison of the energy consumption with conventional flash memories It is difficult to compare directly the energy consumption of our device with conventional flash memories since, first of all, the operation concept of our device especially in the active region (MEMS part) is completely different and second, the conventional flash memory usually receives power with other cells in the same word through a common cell source line, the so-called string. In a conventional flash memory, we have to technically apply the power to all 32 cells belong to a word at the same time even though we only need to operate one single cell. Therefore, the power requirement of a flash memory is usually considered by estimating the power consumed in a word. To compare the power consumption of our device with conventional flash memories, we first estimated the energy required for the MEMS cantilever part and compared that with a conventional memory. Then the power requirement of the CNT-FET was compared with a conventional MOSFET which is used as the main component of a flash memory. According to a recent energy characterization study 36 on the conventional multilevel-cell (MLC) flash memories (Intel MLC NOR 28F256L18 and Samsung KAK38200AM), the energy consumption per word (which contains 4 bytes, i.e. 32 cells in a word) during programming varies from 752 nj to 31194 nj for Intel s memory and from 1.67μJ to 51.57μJ for Samsung s memory. Therefore, the energy consumption of conventional flash memories per cell is estimated to be on the order of 20 to 1000 nj. As we mentioned in the paper, the MEM cantilever is operated via electrostatic actuation and it does not consume any current so that only electrostatic energy and almost zero power is required during the operation. The electrostatic energy can be simply expressed by the following equation for the capacitive energy: C is the capacitance between cantilever and actuation electrode, V is the voltage applied for actuation (10V), and are the relative permittivity and vacuum permittivity, A is the overlap area between cantilever and bottom electrode (2μm 5μm), d is the distance between cantilever and actuating electrode (500nm). The capacitance value estimated from the above parameters was on the order of 100 af. The resultant capacitive
energy of the MEM cantilever during operation is calculated to be 5 10-14 J. Therefore the energy consumption of the cantilever part is almost negligible compared to that of a conventional memory. A recent report (ref. 30) shows that the power consumption of a CNT-FET is comparable to a MOSFET. The power consumption of our working CNT-FET device which was described in this work was estimated to be around 10nW at on-state current (10-7 A with 100mV of bias voltage). The energy consumption of this CNT-FET device is almost comparable to or less than that of a conventional flash memory cell even though the bias voltage is applied continuously to the CNT FET device. Note that the actual reading time of a conventional flash memory is on the order of tens of μs per page. Technically, the energy consumption of our CNT-FET device can be considerably decreased if the time for applying bias to the CNT FET is controlled to be of the same order as the reading time of a conventional flash memory.
Supplementary Reference 36. Joo, Y., Cho, Y., Shin, D., Park, J. & Chang. N. An energy characterization platform for memory devices and energy-aware data compression for multilevel-cell flash memory. ACM Transactions on Design Automation of Electronic Systems. 13, Article 43 (2008).