Quantifying the Noise Penalty of an ADC Driver Amplifier

An ADC driver can provide gain, impedance transformation, and single-ended to differential conversion, but its noise also reduces system SNR. A direct way to quantify that penalty is to express both converter and amplifier noise as rms voltages referred to the ADC input, then compare and combine them over the actual noise bandwidth.

1 Establishing the signal chain and its performance baseline

An ADC generally provides strong dynamic performance when the applied signal approaches its rated full-scale range. Real sources may require a variable gain amplifier, or VGA, to reach that level. Selection must consider input range, signal frequency and bandwidth, resolution or SNR, and spurious-free dynamic range, SFDR. Two individually capable devices do not automatically form a suitable combination.

The example uses an AD8375 intermediate-frequency VGA with an AD9268 dual-channel 16-bit pipeline ADC. The AD9268 is offered at 125, 105, and 80 MSPS and specifies 78.2 dBFS SNR and 88 dBc SFDR at a 70 MHz IF. The AD8375 supports single-ended to differential conversion with good linearity and relatively consistent noise across gain settings. Its gain is set to +6 dB in this example.

A fifth-order Butterworth low-pass antialiasing filter connects the devices. Its nominal bandwidth is about 100 MHz, with 150 Ω input and output impedances. The VGA has a high differential output impedance modeled as 16 kΩ in parallel with 0.8 pF. The measured response in Figure 2 marks the −3 dB frequency at approximately 105 MHz; that value is used in the noise integration below.

Figure 1 Filtering and impedance network between the AD8375 and AD9268

Figure 2 Signal-chain response with a cutoff near 105 MHz

2 Comparing front ends to estimate additional noise

The evaluation board's passive path uses a transformer or balun for single-ended to differential conversion and a low-pass network with a cutoff near 200 MHz. The alternative path adds an amplifier and its antialiasing filter. FFT measurements with a 10 MHz single-tone input provide a basis for estimating the extra noise of the active front end.

Figure 3 Passive coupling and amplifier-driven front ends

The passive configuration measures 77.7 dBFS SNR, compared with 72.5 dBFS for the active configuration, a reduction of 5.2 dB. Because the two front-end networks are not identical, attributing the increase to the amplifier should be checked against its noise density under the actual load.

Figure 4 FFT of the passive front end with a 10 MHz input

Figure 5 FFT of the amplifier front end with a 10 MHz input

Datasheet noise densities are not always directly comparable. ADC values are commonly expressed in dBFS/Hz or dBm/Hz, whereas amplifier values are often given in nV/√Hz. Noise figure, NF, is the input-to-output SNR ratio in decibels and is useful in RF analysis, but reference assumptions can make it misleading in an ADC signal chain. Converting the contributions into rms noise voltages provides a clearer budget.

3 Recovering rms noise from SNR

Start with the converter's specified full-scale differential input and convert the sinusoidal peak-to-peak voltage to rms. The full-scale rms value here is 0.707 V. In the following equations, SNR is referenced to full scale.

VFS,rms = Vpp / (2√2)

SNR = 20 log10(VFS,rms / Vnoise,rms)

Vnoise,rms = VFS,rms / 10^(SNR/20)

Using the passive front-end result of 77.7 dBFS gives a converter baseline of approximately 92.2 μV rms. The active-chain result of 72.5 dBFS gives approximately 168 μV rms total noise. Uncorrelated sources add in power, so their voltage amplitudes must not be subtracted directly.

Vtotal² = VADC² + Vamp²

Vamp = √(168² − 92.2²) ≈ 140 μV rms

The inferred amplifier contribution is about 1.5 times the ADC noise. It is therefore a significant limitation on this configuration's dynamic performance. Adequate drive capability alone does not ensure that system SNR will approach the converter's standalone specification.

4 Checking the result against load and noise bandwidth

The AD8375 specifies approximately 20 nV/√Hz with a 150 Ω differential load. Although its noise changes relatively little with gain, it does depend on loading. The density must therefore be adjusted for the network in Figure 1. With the high output impedance approximation, the effective load is:

Zload ≈ 10 Ω + (300 Ω ∥ 150 Ω ∥ 3.5 kΩ) ≈ 107 Ω

NSD ≈ 20 × 107/150 = 14.3 nV/√Hz

A real low-pass filter passes some noise above its −3 dB frequency. The ratio of equivalent noise bandwidth to −3 dB bandwidth is the shape factor, which approaches unity as filter order increases. Table 1 retains the order-to-factor relationship used in the example.

Table 1 Filter order and noise bandwidth shape factor

The fifth-order filter has a shape factor of 1.02. Multiplying the noise density by the square root of the effective bandwidth, using the 105 MHz response in Figure 2, gives approximately 149 μV rms.

Vamp,calc = NSD × √(BW × shape factor)

Vamp,calc = 14.3 nV/√Hz × √(105 MHz × 1.02) ≈ 149 μV rms

The estimate of 149 μV is reasonably close to the 140 μV inferred from measurement, supporting the conclusion that the amplifier is the principal noise limitation in this example. Exact agreement is not required, but the comparison helps detect errors in loading, bandwidth, or reference units.

5 Including noise in the initial component choice

The benefits of a driver and its noise penalty should be evaluated together. Early in the design, establish the full-scale rms voltage and converter noise, adjust amplifier noise density for the actual load, integrate over the filter's equivalent noise bandwidth, and combine the contributions to predict SNR. This exposes an unsuitable combination before implementation rather than during final testing.