An Information Theoretic Approach to Optimal Amplifier Operation Nicole M. Nelson and Pamela A. Abshire Electrical and Computer Engineering/Institute for Systems Research University of Maryland, College Park MD, 20742 USA nmnelson,pabshire@umd.edu function of input signal power. We compare the theoretical capacity with an empirical capacity computed from measurements of noise and transfer functions. Abstract— Amplifier performance can be severely degraded by the presence of intrinsic physical noise sources. We analyze the information transfer rates for simple and wide-range operational transconductance amplifiers (OTAs) using the principles of information theory. Frequency transfer characteristics and input-referred noise of both simple and wide-range OTAs are modeled using process dependent noise parameters to determine the frequency ranges of operation that provide minimum noise and maximum information capacity. The theoretical predictions of channel capacity are compared with empirical estimates of capacity determined using transfer characteristics and output noise obtained experimentally. The amplifiers have been fabricated in a commercially available 0.5µm process. I. II. Many physical systems can be modeled as Gaussian channels corrupted by noise. Noise is assumed to be contributed additively at the output of the channel, independently of the input signal. The capacity of the information channel is obtained by maximizing the mutual information between the input and output. For a Gaussian channel subject to input signal power constraint P, the capacity increases linearly with bandwidth F and logarithmically with signal-to-noise ratio according to the classical expression INTRODUCTION C = F log 2 (1 + As CMOS devices and supply voltages scale down, fundamental noise limits remain the same causing a decrease in the achievable signal to noise ratio. Noise is therefore a significant source of performance limitation in integrated circuits. A measure of circuit performance may be obtained by applying the methods of information theory to analog circuits. In prior work [1-3], methodology was formulated to determine the efficiency of information transfer rates through analog circuits by modeling the circuits as Gaussian channels and determining the information capacity from circuit models of input-referred noise. The theoretical models were validated by comparison to capacity computed from measured transfer functions and noise. In [1] these methods were applied to a wide-range operational transconductance amplifier (OTA). P N ) bits sec (1) where C is the capacity of the Gaussian channel and N is the noise power. In the case of colored Gaussian noise, the capacity is given by C = max 2 S ( f ):σ < P ∫ log (1 + 2 S( f ) N( f ) ) df (2) where the input signal power spectral density S(f) that maximizes the information transfer and provides capacity is determined using the water-filling technique in which the input signal power is allocated to the frequency bands with lowest noise spectral density, given by: f2 Analysis in [1] modeled the input-referred noise spectral density as a convex function, since only gate source capacitances were considered. In this work, we consider a more general small signal model of the MOSFET, including gate-drain capacitances, in order to accurately determine input-referred noise, optimal operating frequencies, and capacity. We apply algorithms developed in [1] with a more accurate small signal model to determine the input-referred noise of simple and wide-range OTAs and use this information to determine the information capacity as a 0-7803-9197-7/05/$20.00 © 2005 IEEE. GAUSSIAN CHANNEL CAPACITY P = ∫ S ( f ) df = f1 f2 ∫ (υ − N ( f ) ) + df (3) f1 Figure 1 illustrates the technique: the input-referred noise of a simple OTA is shown, along with the signal power allocation (shaded area) that allows the channel to transmit information at the channel capacity. 13 Figure 3. Measured noise of NMOS transistor (W=1750 µm L=7µm) in a commercially available 0.5 µm process. Figure 1. Input-referred noise of simple OTA. Shaded area is the signal allocation that provides maximum information transmission. A. Noise Sources in a MOS Transistor The four main sources of noise in a MOS transistor include: thermal noise due to the thermal excitation of charge carriers in a conductor; flicker noise due to electron traps caused by imperfections in the silicon-silicon dioxide interface; shot noise generated by the channel current; and gate current noise caused by thermal fluctuations in the channel. Gate current noise is negligible, and shot noise is caused by the same mechanism as thermal noise [3]. Therefore we consider only flicker and thermal noise. Both components are referred to the input of the transistor (the gate) by dividing by the gain, giving: (a) Vi 2 ∆f 4 kT γ gm + Kf ID Af g m 2Cox L2 f β (4) where kT is the thermal energy, L is the device length, Cox is the oxide capacitance, ID is the channel current, gm is the transconductance, and Kf , Af and β are process dependent constants. γ is a constant equal to 1/2 for subthreshold and 2/3 for above threshold operation. The process dependent parameters Kf , Af and β were extracted according to experimental procedures developed in [4,5] using an Agilent 4395A Network/Spectrum/Impedance Analyzer. Equation (4) can be rewritten as (b) Figure 2. a) Simple OTA. b) Wide-Range OTA. log III. = ( ) = log ( Vi 2 ∆f K f ID Af Cox Leff 2 gm 2 ) − β log ( f ) (5) The noise spectral density measured for an NMOS transistor of size W =1750 µm and L=7µm is shown in Figure 3. This noise data has been fitted using least-squares methods to determine noise parameters Kf=10-24.49V2/F and β=1. Since fairly large changes in the drain current are necessary to produce measurable change in the spectral density for the chosen sizes, Af was assumed to be 1. INPUT-REFERRED NOISE MODEL Circuit diagrams for the simple and wide-range operational transconductance amplifiers under consideration are shown in Figure 2 (a) and (b), respectively. The transfer functions and input-referred noise for both simple and widerange OTAs are modeled and used to determine the information capacity. Both circuits are considered to be composed of a cascade of analog transistors each with its own transfer function and equivalent input noise sources. 14 Figure 4. Calculated input-referred noise for the simple and widerange OTAs at several bias currents. Figure 5. Measured input-referred noise for the wide-range OTA at bias current of 10µA. B. Wide-Range OTA The input-referred noise of the wide-range OTA (Figure 2b) was derived in the same manner as the simple OTA above assuming balanced transconductances gm1=gm2, gm3=gm4=gm5=gm6 and gm7=gm8 and equal output resistances. For a wide-range OTA fabricated in the same 0.5µm process, these transfer characteristics yield a low pass function with a cut-off frequency around 10KHz and open loop gain of 40 at low frequencies. Figures 4 and 5 show the theoretical and measured input-referred noise of the wide-range OTA as a function of frequency, respectively for several values of the bias current (Ibias=0.1, 1, 10µA) in Figure 4 and for Ibias=10µA in Figure 5. The information capacity was calculated as a function of average input signal power via waterfilling and is shown in Figure 6 below. B. Effect of Gain and Feedback For a system composed of i analog blocks each with its own noise source Vni and gain Ai in each stage the inputreferred noise is given by 2 Vin 2 = Vn12 + VAn 22 + 1 Vn 32 A12 A22 + ... + Vni 2 A12 A2 2 ... Ai−12 (6) For ideal feedback the equivalent input noise generators may be moved unchanged outside the feedback loop without affecting circuit noise performance CALCULATED RESULTS IV. A. Simple OTA The small signal transfer function for the simple OTA shown in Figure 2a was determined using a small signal model that includes the effects of both gate-drain and gatesource capacitances. All calculations assume balanced transconductances gm4=gm1 and gm3=gm2 and equal output resistances. The transfer function is given by: Vo1 V+ = Vo 2 V− = ( sC ( sCgd 1 − gm1 )( sCgd 2 − gm 2 ) 1 gd 2 + ro )( sC 2 gd 1 + ro + g m 3 + 2 sC gs 2 For both OTAs the input-referred noise is not a simple convex function as reported by [1]. The total input-referred noise depends on the transfer function which in turn depends on the MOSFET parasitic capacitances. Considering only gate-source capacitances leads to the erroneous conclusion that the noise increases without bound as a function of frequency. Including the gate-drain capacitances does not significantly change the lower frequency noise spectrum, but at high frequencies the Miller effect causes the slope of the transfer function to decrease, which in turn causes the noise spectral density to deviate from the expected f2 slope at high frequencies. In fact, the model predicts that the noise spectrum approaches an asymptotic value as frequency increases (see Figure 4). ) ( sCgd 1 − gm1 ) ( sC 1 gd 1 + ro ) (7) Vo = Vo1 + Vo 2 This means that the input-referred noise is no longer cup shaped, so once the dip in the input-referred noise spectrum has been completely filled with input signal spectral power, further increases in the input signal power cause the capacity to increase as if the noise were spectrally uniform. For white noise the capacity increases linearly with the signal power. The minimum value of the input-referred noise spectral density shifts to higher frequencies as the bias current increases, as shown in Figure 4. This minimum noise which, for bias current Ibias=10uA, is a low pass function with cut off frequency around 10KHz and open loop gain of 20 at low frequencies. Figures 1 and 4 show the measured and theoretical input-referred noise of the simple OTA as a function of frequency, respectively, at the bias current Ibias=10µA. Water filling is performed on the input-referred noise spectral density to compute its capacity as a function of bias current and input signal power. 15 wide-range OTA uses PMOS transistors while the simple OTA uses NMOS input transistors. For an input signal power of 1µW the simple OTA has a theoretical capacity of 3.36×108 bits/sec while the wide-range OTA has a theoretical capacity of 2.06×107 bits/sec, which matches the empirical capacity of 1.45×107 bits/sec and [1] which calculated a theoretical capacity of 3.58 × 107bits/sec. Therefore the new model provides a tighter estimate of capacity, and we conclude that the parasitic gate-drain capacitances play an important role in shaping input-referred noise and determining capacity of the overall circuit. VI. We analyzed the information transfer rates for simple and wide-range operational transconductance amplifiers using the principles of information theory. Frequency transfer characteristics and input-referred noise of both simple and wide-range OTAs are modeled and used to determine the frequency ranges of operation that provide minimum noise and maximum information. Theoretical predictions of channel capacity are compared with empirical estimates of capacity determined using measured transfer characteristics and output noise. The models in the present work improve upon previously reported work [1] in that they account for gate-drain capacitances. This allows for tighter bounds on the capacity which agree very well with empirical estimates. Figure 6. Calculated capacity vs signal power for wide-range OTA. frequency depends on the process dependent parameters Kf and Af. Lower values of Kf can cause the noise spectral density to exhibit a peak between the minimum and the asymptotic regime, an interesting observation which will be investigated further in future work. The larger the value of Af is, the higher the frequency at which the minimum noise occurs. Since Kf and Af are process dependent parameters, the fabrication technology plays a crucial role in determining noise and capacity. V. CONCLUSIONS EXPERIMENTAL RESULTS ACKNOWLEDGMENT A chip with implementations of the simple and widerange OTAs shown in Figure 2 was fabricated in a commercially available 0.5 µm process and used to determine the transfer characteristics and output noise. The output noise spectral density for an NMOS transistor shown in Figure 3 was measured using an Agilent 4395A Network/Spectrum/Impedance Analyzer in order to derive process-dependent noise parameters Kf and β. The same instrument was used to measure output noise and frequency transfer characteristics for simple and wide-range OTAs for several different bias currents (not shown). From these measurements the input-referred noise spectral densities shown in Figures 1 and 6 were computed. Finally the empirical capacities for the simple and wide-range OTAs were estimated from the computed input-referred noise. We thank the MOSIS service for chip fabrication. This material is based on work supported by the Laboratory for Physical Sciences and by the National Science Foundation (P.A.CAREER Award 0238061). REFERENCES [1] [2] [3] [4] Figure 6 shows the information capacity for a wide-range OTA computed in three ways: using the model in [1], using the analysis and model in this paper and from experimental data reported in this paper. The new model results in lower values of capacity, so it provides a tighter bound on the information transmission of the circuit. In fact at low values of signal power the new model agrees extremely well with capacity computed from experiment. [5] [6] [7] The optimal frequency range of operation for the simple OTA is in the 10 to 100kHz range while for the wide-range OTA the optimal frequency band is lower, in the 1 to 10 kHz range. At low frequencies, measured output noise of the simple OTA is higher than for the wide range OTA. This result is as expected, since the input differential pair of the [8] 16 M Loganathan, S Malhotra, P Abshire, “Information Capacity and Power Efficiency in Operational Transconductance Amplifiers,” Proc ISCAS , vol 1. pp.193–196, May 2004. Furth, P. and Andreou, A.G., “A Design Framework for Low Power Analog Filter Banks,” IEEE TCAS I, vol. 42(11), pp. 966-971, November 1995. Abshire, P. and Andreou, A.G., “Capacity and Energy Cost of Information in Biological and Silicon Photoreceptors,” Proc IEEE, vol. 89, pp. 1052-1064, July 2001. R Sarpeshkar, T Delbruck, CA Mead, “White noise in MOS transistors and resistors,” IEEE Circuits Devices Mag, pp. 23 – 29 Nov 1993. Y Nemirovsky, I Brouk, C Jakobson, “1/f Noise in CMOS Transistors for Analog Applications,” IEEE Trans on Electron Devices, vol 48, No 5, pp. 921 – 927 , May 2001. A Blaum et al., “A New Robust On-Wafer 1/f Noise Measureemnet and Characterization System,” Proc. IEEE 2001 Intl. Conf. on Microelectronic Test Structures, pp. 125-130, March 2001. TM Cover and JA Thomas, Elements of Information Theory, John Wiley & Sons Inc., New York, 1991. RR Harrison and C Charles, “A Low-Power Low-Noise CMOS Amplifier for Neural Recording Applications,” IEEE JSSC, vol. 38, pp. 958 – 965, 2003.