A supply voltage (VDD) independent temperature sensor circuit, which can be realized by the optimum combination of three current modes of n-MOSFETs including the subthreshold current using the feedback scheme from the temperature dependent voltage (VTD) output to the gates of three n-MOSFETs, was proposed and fabricated by a standard 1.2 µm n-well CMOS process. The circuit consists of only 17 MOSFETs without high resistors resulting in a small die area of 0.18 mm2. The temperature coefficient TC of the sensor circuit can be controlled by the channel length ratio L4/L3 of two n-MOSFETs. The average temperature sensor voltage VTS and its typical TC are 1.77 V at VDD=5.0 V (20) and 5.1 mV/ for VDD=5.01.0 V in the temperature range of -20-100 in case of L4/L3=9, respectively.

In this work, the metal-oxide-metal (MOM) capacitor in the scaled CMOS process has been modeled at high frequencies using an EM simulator, and its layout has been optimized. The modeled parasitic resistance consists of four components, and the modeled parasitic inductance consists of the comb inductance and many mutual inductances. Each component of the parasitic resistance and inductance show different degrees of dependence on the finger length and on the number of fingers. The substrate network parameters also have optimum points. As such, the geometric dependence of the characteristics of the MOM capacitor is investigated and the optimum layout in the constant-capacitance case is proposed by calculating the results of the model. The proposed MOM capacitor structures for 50fF at f =60GHz are L =5μm with M =3, and, L =2μm with M =5 and that for 100fF at f =30GHz are L =9μm with M =3, and L =4μm with M =5. The target process is 65-nm CMOS.

A test structure to separately measure sheet resistances of highly-doped-drain (HDD) and lightly-doped-drain (LDD) in LDD-type CMOSFETs with various gate spaces S having sub-100 nm sidewalls was proposed. From the reciprocal of source/drain-resistance R-1 versus S characteristics, the sheet resistance ρH of the high-conductive-region (HCR) corresponding to HDD and the approximate width WLC of the low-conductive-region (LCR) corresponding to LDD could be estimated. Both of ρH and WLC for p- and n-MOS devices were scarcely dependent on the gate voltage. The sidewall-width difference of 40 nm could be sufficiently detected by using the test structure with the S pitch of about 60 nm. The R-1 versus S characteristics showed the unstable resistance variations in the narrow S region less than 0.3 µm, which corresponded to the minimum S for the process used for the test device fabrication and suggested that various micro-loading effects seriously affected on the characteristics.

A test structure to separately analyze the location where the hot-carrier-induced CMOSFET reliability is determined around the center or the isolation-edge along the channel-width was proposed and fabricated. The test structure has four kinds of MOSFETs; [A] and [D] MOSFETs with a short and a long channel-length all over the channel width, respectively, [B] MOSFET with the short and the long channel-length around the center and the both isolation-edges, respectively, and [C] MOSFET with the channel-length regions vice versa to the [B] MOSFET. The time dependent changes of the threshold voltages VT, the saturation currents IS, the linear currents IL and the maximum transconductances β up to 50,000 s were measured. All data for the wide channel-width MOSFETs were almost categorized into three; [A], [B]/[C] and [D]. The [B]/[C] data were well estimated from simple theoretical discussions by the combination of [A] and [D] data, which mean that the reliabilities are nearly the same around the center or the isolation-edge for the CMOSFETs.

We have developed a digital coprocessor with a dynamically reconfigurable pipeline architecture specified for a layered neural network which executes on-chip learning. The coprocessor attains a learning speed of 18 MCUPS that is approximately twenty times that of the conventional DSP. This coprocessor obtains expansibility in the calculation through a larger multi-layer, network by means of a network decomposition and a distributed processing approach.

Processing elements (PEs) with a dynamically reconfigurable pipeline architecture allow the high-speed calculation of widely used neural model which is multi-layer perceptrons with the backpropagation (BP) learning rule. Its architecture that was proposed for a single chip is extended to multiprocessors' structure. Each PE holds an element of the synaptic weight matrix and the input vector. Multi-local buses, a swapping mechanism of the weight matrix and the input vector, and transfer commands between processor elements allow the implementation of neural networks larger than the physical PE array. Estimated peak performance by the measurement of single processor element is 21.2 MCPS in the evaluation phase and 8.0 MCUPS during the learning phase at a clock frequency of 50 MHz. In the model, multi-layer perceptrons with 768 neurons and 131072 synapses are trained by a BP learning rule. It corresponds to 1357 MCPS and 512 MCUPS with 64 processor elements and 32 neurons in each PE.

MOS capacitors with Si-implanted thermal oxide and CVD deposited oxide of 30 nm thickness were fabricated for applications of non-volatile memory and electroluminescence devices. Current-voltage (I-V) and I-V hysteresis characteristics were measured, and the hysteresis window (HW) and the integrated charge of HW (ICHW) extracted from the hysteresis data were discussed. The HW characteristics of high Si dose samples showed the asymmetrical double-peaks curves with the hump in both tails. The ICHW almost converged after the 4th cycle and had the voltage sweep speed dependence. All +ICHW and -ICHW characteristics were closely related to the static (+I)-(+VG) and (-I)-(-VG) curves, respectively. For the high Si dose samples, the clear hump currents in the static I-VG characteristics contribute to lower the rising voltage and to steepen the ICHW increase, which correspond to the large stored charge in the oxide.

We have developed a 64-neuron electrically trainable BiCMOS analog neuroprocessor based on 3-layered PDP networks with a feedforward time as short as 10 µs which is equivalent to the operation speed as high as 108 multiplications per second. A crucial point in this development is application of a dynamic refreshment technique to a weighting circuit. A sufficiently long retention time of the synapse weight has thereby been attained, leading to a practical operation of the neuroprocessor.

A striped inductor and its utilization of a voltage-controlled oscillator (VCO) are studied with the aim of suppressing phase noise degradation in K- and Ka-bands. The proposed striped inductor exhibits reduced series resistance in the high frequency region by increasing the cross-sectional peripheral length, as with the Litz wire, and the VCO of the striped inductor simultaneously exhibits a lower phase noise than that of the conventional inductor. Striped and conventional inductors and VCOs are designed and fabricated, and their use of K- and Ka-bands is measured. Results show that the Q factor and corner frequency of the striped inductor are approximately 1.3 and 1.6 times higher, respectively, than that of the conventional inductor. Moreover, the 1-MHz-offset phase noise of the striped inductor's VCO in the K- and Ka-bands was approximately 3.5 dB lower than that of the conventional inductor. In this study, a 65-nm standard CMOS process was used.

A test structure to analyze asymmetry and orientation dependence of MOSFETs is presented. n-MOSFETs with 8 different channel orientation and three kinds of process conditions were measured and symmetry characteristics of IDsat and IBmax with respect to the interchange of source and drain was examined. Although both IDsat and IBmax have similar channel orientation dependence, IBmax in interchanged S/D measurements shows asymmetrical characteristics, which can be applied to a sensitive method for device asymmetry detection.