Why Voltage-Reference Noise Matters in High-Resolution ADC Systems
In converters above 24 bits, the ADC itself is often no longer the only noise bottleneck. The voltage reference is part of the conversion ratio, so reference noise is multiplied into the output code as the input signal approaches full scale.
The source provides a useful scale: achieving roughly 25-bit resolution, or about 152 dB of dynamic range with a 5 V full-scale signal, allows only about 0.2437 µV rms of total system noise. At that level, the reference, buffer, filter, capacitor ESR, and even leakage must all be treated as signal-chain components.
Relationship between noise, ENOB, and SNR
1. Why Shorted-Input ADC Tests Hide Reference Noise
ADC data sheets often characterize noise with the analog input shorted. When VIN = 0, the VIN/VREF term is also zero, so reference noise does not translate into output-code variation. This is useful for measuring the converter's intrinsic noise, but it does not represent full-scale operation.
ADC Code = VIN × 2^N / VREF
As VIN increases, reference-voltage fluctuations increasingly appear as gain error. In the source test using AD7177-2, LTC6655-5, and a low-noise dc source, total rms system noise rises with the ADC input level.
RMS system noise versus ADC input voltage
2. A Large Output Capacitor Is Not Always the Best Answer
A large reservoir capacitor at the reference output serves two purposes: it forms a low-pass network and supplies transient current to the ADC reference input. However, the reference output stage is itself a feedback amplifier, and a large capacitive load changes loop phase margin.
Passive reference-output network including output resistance, ESR, and storage capacitor
fC = 1 / (2πRC)
Some SAR ADCs require at least about 10 µF on the REF pin, but that load can create noise peaking near the reference-buffer unity-gain crossover. ESR can sometimes damp the main peak, but it can also introduce additional high-frequency behavior. The correct capacitor value and ESR must therefore be verified against the specific reference design.
3. A More Robust Architecture: Filter First, Buffer Second
A common improvement is to place a low-frequency RC filter after the reference and then use a low-noise buffer to drive the ADC. This separates two conflicting tasks: narrow-band noise reduction and high-speed transient drive.
Passive RC low-pass followed by a reference buffer
The RC pole attenuates wideband and low-frequency reference noise, while the buffer supplies the dynamic current demanded by SAR or sigma-delta ADC reference pins. The buffer must combine low noise, capacitive-load stability, low distortion, adequate slew rate, and sufficient gain bandwidth.
4. When a Higher-Order Active Filter Becomes Useful
If the required ENOB is very high, a single 20 dB/decade pole may not provide enough broadband-noise attenuation. A higher-order active low-pass filter can deliver a much steeper roll-off.
Examples of Bessel, Butterworth, Chebyshev, and elliptic low-pass responses
For a voltage-reference path, a flat passband is often more important than aggressive passband ripple, making Butterworth behavior attractive. The source also discusses a signal-flow-graph active RC method that allows independent control of cutoff frequency and Q, at the cost of more power, area, and BOM.
Active low-pass implementation derived from a signal-flow graph
5. Integrated Noise-Reduction Pins Move the Filter Ahead of the Buffer
A reference with an integrated noise-reduction node allows the RC filter to operate ahead of the output buffer. This can preserve low output noise without sacrificing transient drive capability. The source uses LTC6655LN as an example.
LTC6655LN architecture with external noise-reduction capacitor
Leakage becomes critical because the NR pin is a sensitive, high-impedance node. Even small leakage currents can create ppm-level dc error.
6. System Performance Is Shared by the ADC, Reference, and Input Source
The source evaluates multiple reference configurations with a 32-bit AD7177-2. At higher output data rates the converter's own noise dominates; at lower data rates, flicker noise from the reference and dc source becomes increasingly important.
Noise-density comparison of system contributors
ENOB versus ADC output data rate
The measurements show substantial broadband-noise reduction when the LTC6655LN noise-reduction capacitor is used, while the higher-order SFG filter provides even stronger attenuation and can improve achievable ENOB at low output data rates.
7. Practical Design Flow
1. Translate the target ENOB, full-scale voltage, and bandwidth into an allowable total rms-noise budget.
2. Do not rely only on shorted-input ADC noise; evaluate reference noise near full-scale input.
3. Before placing a large capacitor directly on a reference output, check capacitive-load and ESR stability guidance.
4. For very low cutoff frequencies, use an RC filter followed by a buffer or a reference with a dedicated NR node.
5. When using higher-order active filtering, include op-amp noise, drift, power, and BOM in the total budget.
6. Validate ENOB and SNR with the real ADC, real reference, and real input source.
Conclusion
In a high-resolution converter, the reference is not an ideal numerical constant; it is an analog signal path that directly sets the conversion scale. Reference noise, output impedance, capacitive-load stability, ESR, filter noise, and buffer transient performance all influence the final ENOB. Converting ADC resolution into real system accuracy requires the reference path to be engineered as carefully as the front-end amplifier.