No abstract

This paper presents a 4-tap coefficient-error-robust feed-forward equalization (FFE) transmitter (TX) for massively parallel links. Recently, massively parallel links such as on-chip links [1][2][3], silicon interposers [4,5], or wide I/Os [6] are gaining popularity to meet increasing demand for data transmission with a limited power budget. However, calibration overhead for thousands I/Os to compensate coefficient errors due to nano-scale variation has a high hardware cost. To reduce this overhead, we develop a coefficient-error-robust FFE (B-FFE) TX architecture that uses the channel loss to suppress eye perturbation due to coefficient errors while behaving identically to a conventional FFE.Without coefficient errors, a B-FFE TX can be designed identical to any FFE TX. Figure 2.7.1 depicts the architectures of the B-FFE and a conventional FFE. A B-FFE TX contains a digital transition-detection (TD) filter that detects transitions of incoming data and generates transition signal: '1' for '-1'→'1' data transition; '-1' for '1'→'-1' data transition; and '0' for no transition. The transition signal is delayed by a chain of 1UI delay units. Each delayed transition signal is weighted by a coefficient (a k , k ≠ 0) and added to the incoming data weighted by a 0 , to generate the output voltage. If designers configure the FFE coefficients (w k ) and the B-FFE coefficients (a k ) as a 0 = ∑ N-1 w i and a k≠0 = -2∑ N-1 w i then the output voltages of both FFE and B-FFE TXs are identical.Although the nominal behaviors of FFE and B-FFE are identical, the coefficienterror-tolerance of B-FFE is superior to FFEs. Figure 2.7.2 explains how B-FFE suppresses signal perturbation caused by a coefficient error. Typical coefficient errors due to mismatch, process-temperature variation, or supply-voltage drop can be modeled as an additive constant (Δw k and Δa k ) to the nominal coefficient (w k and a k ) as Fig. 2.7.1 shows. Since the FFE and B-FFE TXs are linear timeinvariant systems, a coefficient error perturbs the pulse response of the TX by adding an error pulse, which is the pulse at the tap-position multiplied by the coefficient error as shown in Fig. 2.7.2. For the FFE TX, the error pulse at TX is square-shaped having large low-frequency portion, and thus is not significantly attenuated by the low pass filter (LPF) channel (-25dB loss at Nyquist frequency 4GHz). However, the error pulse of the B-FFE TX except the first tap (Δa 0 ) has large high-frequency portion after the TD modulation. Therefore, the LPF channel significantly attenuates the B-FFE error pulse, and thus, at the receiver, the impact of the B-FFE coefficient error is significantly attenuated, improving tolerance to the coefficient errors. The error caused by the first tap (Δa 0 ) of a B-FFE is typically insignificant since the first tap size, which determines the DC level, is typically small for a lossy channel.To comparatively analyze the concept, we design a test-chip containing a 4-tap B-FFE TX and a 4-tap FFE TX. Figure 2.7.3 depicts the simplified block d...

Abstract-This paper proposes a cost-efficient and automatic method for large data acquisition from a test chip without expensive equipment to characterize random process variation in an integrated circuit. Our method requires only a test chip, a personal computer, a cheap digital-to-analog converter, a controller and multimeters, and thus large volume measurement can be performed on an office desk at low cost. To demonstrate the proposed method, we designed a test chip with a current model logic driver and an array of 128 current mirrors that mimic the random process variation of the driver's tail current mirror. Using our method, we characterized the random process variation of the driver's voltage due to the random process variation on the driver's tail current mirror from large volume measurement data. The statistical characteristics of the driver's output voltage calculated from the measured data are compared with Monte Carlo simulation. The difference between the measured and the simulated averages and standard deviations are less than 20% showing that we can easily characterize the random process variation at low cost by using our costefficient automatic large data acquisition method. Index Terms-Random process variation, costefficient measurement, automatic large volume data acquisition, statistical characterization, current mode logic driver, current mirror array

No abstract

In this research, an one selector-one ReRAM (1S1R) cross-point array of a multi-level cell (MLC) was demonstrated and investigated. To expand high-density feasibility of cross-point array, MLC pulse writing and reading operations were assessed with parasitic line resistances and capacitances using Matlab and HSPICE simulations. We observed a switching energy is an important parameter for MLC in actual cross-point array in the operating point of view. In addition, not only ReRAM but also selector characteristics are highly important in the device point of view. Therefore, this study serves power efficient guidelines for 1S1R devices and operating schemes of cross-point array.