In recent years, there has been an increasing interest in the application of portable battery powered biomedical devices in consumer products. To increase the operation time of the battery employed in biomedical products, the system needs to be designed for low-voltage and low-power consumption. Several biomedical applications like electro-cardiogram (ECG), electroencephalogram (EEG), and evoked potential (EP) employ digital signal processing (DSP) techniques after digitizing the sensed analog signals. The analog-to-digital data converter (ADC), used for digitizing the sensed analog signal, is a critical and power hungry block in a generic biomedical system [1].

The ADC employed in biomedical signal processing chain typically requires 1–6 kHz signal bandwidth with up to 14-bits or higher resolution [1]. Delta-sigma modulation based ADCs employing an oversampling technique are a well-known candidate for achieving higher resolution analog-to-digital conversion using a low-resolution quantizer [2,3,4,5]. Using oversampling techniques, the quantization noise introduced by the low-resolution quantizer is pushed outside the signal bandwidth, which results in higher resolution (SNR) and dynamic range (DR). Also, ΔΣ ADCs are well-suited for digitizing biomedical signals without requiring any instrumentation amplifier. When compared to other conventional ADCs, small differential-mode signal can be recovered with high resolution using a ΔΣ modulator [2,3,4,5].

The ΔΣ ADC is also proven to be well suitable for low-bandwidth signal conversion by employing oversampling, and is robust to device mismatches in the forward signal path when using nano-CMOS processes [5,6,7]. Traditionally, discrete-time (DT) ΔΣ ADC were the commonly used ΔΣ architecture. Many such modulators have been reported over previous decades [8,9]. However, in recent years, the continuous-time delta-sigma (CT-ΔΣ) ADC is being preferred in low-power signal-processing applications due to several attractive features like an inherent anti-aliasing filter (AAF), relaxed bandwidth requirements of the active elements and lower power consumption when compared to their discrete-time counterparts [3,5,7].

To enhance the power reduction in battery-powered, low-bandwidth biomedical applications, a single-bit CT-ΔΣ ADC is an appropriate choice. Advantages of using a single-bit quantizer include reduced chip area and significant power reduction of the ΔΣ loop. However, some important concerns in designing single-bit CT-ΔΣ include the higher in-band quantization noise due to the use of a single-bit quantizer, increased sensitivity to clock jitter and the slew rate requirement for the first operational amplifier (opamp). The higher in-band quantization noise can be addressed by increasing the effective oversampling ratio (OSR). Since the signal bandwidth requirement of battery powered biomedical applications is low, even with lower sampling rate a higher OSR can be employed. Also, due to the low sampling frequencies used in the CT-ΔΣ ADCs for biomedical applications, clock jitter is not a serious concern.

Any amount of slewing in the opamp introduces distortion and increases the in-band noise, resulting in a reduction of signal to quantization noise (SQNR) [10]. This should be taken care of by careful design of the opamp. The main bottleneck in achieving low-power CT-ΔΣ ADC design is the power requirement of the loop-filter integrators. Any power reduction in opamp reduces the finite DC gain and unity-gain frequency (f_un) of the opamp. The lower f_un of opamp impairs the characteristics of the loop-filter and thus modifies the modulator noise transfer function (NTF) with significant performance degradation. In order to achieve a low-power design with acceptable performance, the systematic design centering method proposed in [11] is adopted in this work. Further, the use of a subthreshold optimized FDSOI CMOS process in this design resulted in a low-power circuit-level implementation.

This invited paper presents an expanded and detailed discussion on our previous work reported in [5]. The remainder of this paper is organized as follows. Section 2 illustrates the modulator architecture for this application. Section 3 demonstrates the design of the third-order modulator using systematic design centering to overcome the opamp non-idealities. Section 4 briefly explains the design of the CT-ΔΣ blocks. Section 5 discusses the simulation results of the designed third-order low power CT-ΔΣ. Finally, Section 6 draws conclusions about the work.

Figure 1 shows the generalized single-ended equivalent block diagram of a single-loop CT-ΔΣ modulator. The ΔΣ ADC consists of an analog modulator followed by a digital decimation filter. In this figure, L(s) is the continuous-time loop filter whose output is sampled and quantized at using a quantizer operating at a sampling rate, f_s, equivalent time period (T_s). The quantizer can either be a single-bit (a comparator) or a multi-bit implementation using a Flash ADC. The modulator employs an oversampling ratio denoted by OSR and the quantizer has an excess-delay of T_d, which is modeled as e^−sTd in Figure 1. Here, u(t) is the modulator input signal, y_c(t) and y[n] are the continuous-time and sampled loop-filter outputs respectively, v[n] is the digital output of the modulator, D_out[n] is the decimated digital output of the ADC with a clock-rate of f_s/OSR. In this section, we describe the various architectural choices made in our CT-ΔΣ modulator design.

Using high-level modeling methods and tools, a loop-filter transfer function is designed using active-RC integrators. Initial modeling uses behavioral models of opamps with suitable parameters for DC gain, unity-gain bandwidth and linearity, and corresponding component values are derived for desired SNDR and dynamic range. After appropriate architecture selection of the opamps, the behavioral opamps are replaced by their transistor-level schematics.

Conventionally, the third-order loop-filter coefficients K= [k₀ k₁ k₂ k₃], are obtained by fitting the impulse response of the discrete-time loop filter L_d(z) = 1 − NTF⁻¹(z) to the continuous-time loop-filter, L_c(s), using the impulse invariance transformation (IIT) for a given feedback DAC pulse shape [14]. The assumption in this method is that the integrators are ideal, with a transfer function given by . This impulse response fitting method breaks down when applied to real integrators with finite op-amp DC gain, non-dominant opamp poles and the presence of opamp zeros. To get the desired NTF_eff(z), either the filter coefficients (K) need to be further tuned or the DC gain and the unity-gain frequency (f_un) of the opamp must be sufficiently large. The latter will result in an increase in the modulator power consumption [11,16]. Also, in a single-bit single-loop CT-ΔΣ modulator, more than 90% of the power is consumed by the opamp, which is a key design block in the loop-filter. In order to achieve low power consumption in the modulator, the bias currents in the opamps need to be optimized for the given specification. Any power reduction in the opamp impairs the characteristics of the loop-filter and thus impacts the resulting NTF performance. Figure 5 shows the magnitude response of NTF(z) with ideal and real integrators for the same loop-filter. It can be seen that the systematic design method restores the modulator NTF with integrators implemented using real opamps having two poles and a zero. Moreover, it is also essential that the slewing in the opamps is sufficiently limited so as not to degrade the overall modulator linearity.

A systematic design centering method for obtaining the loop-filter coefficients by including the effect of the integrator non-idealities is employed [11]. We briefly describe this method for third-order CT-ΔΣ as follows: as we know, the relationship between the noise-transfer function, NTF(z), and the discrete-time loop-filter response, L_d(z), is given

This relation (Equation 1) is incorporated into the impulse response fitting by using the following relation [5]

where h[n] is the impulse response of the NTF(z). Further , where and represent the sampled DAC pulse response of direct path and at the output i^th integrator (see [11] for details). Table 1 shows the derived loop-filter coefficients for the modulator, before and after systematic design procedure.

As described earlier in this section, the first integrator plays a dominant role in determining the desired SNR. The first opamp (A₁) is designed for a DC gain of 75 dB and a unity-gain frequency of f_un = 1.5f_s ≈ 10 MHz with higher slewing performance (≈20 V/μs) [3]. These values are obtained from system-level simulation in our custom developed CT- toolbox (implemented in MATLAB), using a two-pole and single-zero opamp model. Design constraints of the second and third opamp (A₂ and A₃) are relaxed and are designed for f_un = 0.75f_s ≈ 4.5 MHz to consume as low power as possible. The systematic design procedure assumes the adder is ideal. In reality, the adder also introduces excess delay in the loop, due to the finite f_un of the opamp. This excess delay in turn introduces peaking in the magnitude response of the NTF and thus increasing the OBG. To mitigate this NTF peaking, A₄ is designed for f_un = 3f_s ≈ 19 MHz. The power consumption of A₄ is almost equal to that of A₁ as only 60 dB DC gain is used in A₄.

In this section, we describe the circuit level blocks in the CT-ΔΣ modulator. As mentioned earlier, an experimental low-power 0.15 μm FDSOI CMOS process, optimized for sub-threshold (sub-V_T) operation was used for implementing the modulator. The FDSOI process provided a higher ratio for a given transistor overdrive in the sub-V_T regime, and thus lead to lower-power analog circuits when compared to a bulk CMOS process. However, due to the experimental nature of the FDSOI process technology, only typical corner models were available for circuit simulation and the noise parameters were not incorporated in the BSIMSOI Spectre models.

The third-order CT-ΔΣ ADC was designed and fabricated in MIT Lincoln Lab’s experimental 0.15 μm FD-SOI CMOS process. Figure 9 shows the micrograph of the fabricated chip. Since the FDSOI process did not have well-controlled resistivity and capacitance, three separate ADC layouts were included on the chip. The three designs covered a spread of 50% variation in the RC time-constants. Sixty-four thousand output samples were collected from the modulator in the post-layout simulation. Figure 10 and Figure 11 illustrate the post-layout simulated signal to noise and distortion ratio (SNDR). The SNDR is measured using a 2.3 kHz input tone. The peak SNDR is close to 92.5 dB and the simulated dynamic range of the modulator is 94.4 dB. Table 2 summarizes the simulated performance of the modulator. The figure of merit (FoM) of the converter is calculated using power consumption (P), signal bandwidth (f_B) and SNR by using the relation [7].

The modulator presented in this paper achieves 0.271 pJ/level. Table 3 compares the simulated performance of the presented FDSOI design with several continuous-time ΔΣ ADCs from recent literature [7,18,19,20,21,22].

The chip was wire-bonded in a DIP package, mounted on a PCB and tested. External sinusoidal and clock inputs were applied to the chip and the output samples were collected using a mixed-signal oscilloscope. The experimental testing revealed that the value of resistances and capacitances on the chip deviated significantly from the expected values, resulting in the CT-ΔΣ modulators to be unstable. Our experiments led to the conclusion that at least one of the opamps in the design became unstable due to the compensation scheme breaking down with large process variation. Thus a well-characterized process with stable resistivity and capacitance parameters is necessary for such mixed-signal designs. The design presented in this paper shows significant improvement over several design except for an excellent power-optimized CT-ΔΣ design presented in [7].

A low-power, single-bit CT-ΔΣ modulator was designed in a 0.15 μm FD-SOI CMOS process for data conversion employed in generic biomedical applications. The CT loop-filter coefficients were systematically obtained by incorporating the op-amp non-idealities. This method resulted in robust modulator NTF and lower static currents in the opamp. The post-layout simulation results of the CTM exhibit a peak SNDR of 92.4 dB, a dynamic range of 94.4 dB with a maximum stable input (MSA) of −1.92 dBFS. The modulator dissipates 110 μW power from a 1.5 V supply and achieves a FoM of 0.271 pJ/bit. Future work entails fabrication of the chip in a more stable process, investigation of methods for further reducing power in the opamps in the modulator using true subthreshold design, and adaptively tuning to the mitigate the effects of RC time-constant variation in the CMOS process.