Low-noise design techniques
Akos Szoboszlay
© 2014 Akos Szoboszlay
Achieving a low-noise, precision analog design requires using many techniques at different design levels and various design tasks:
• System grounding and isolation
System grounding and related isolation should be designed early — before schematics. The basics are:
• The most sensitive part of the system — usually sensor(s) — should be the ground reference for the entire system, or subsystem containing the sensors.
• Tie reference ground to system ground or earth ground at only one point. This connection should have practically no current flowing through it.
• Draw a block level diagram showing cables to make sure grounding is only at one point, to avoid ground loops.
• Use isolating type DC-DC power supplies.
• Use of digital signal isolators may be required to minimize ground loops and conducted noise.
• Schematic and components:
This is an obvious task for designing the circuit board, but only one of many. Techniques include using low noise op amps, lower value resistors and plenty of bypassing ceramic capacitors (usually 0.1 uF in size 0402), but some wide-bandwidth op amps also require adjacent 10 uF caps (usually use size 0805).
• Layout of circuit board:
Keep noisy components away from noise-sensitive ones both within a circuit board assembly, and also consider nearby noise source(s). Power supplies, solenoid-valves, and transformers can introduce magnetic noise which is harder to shield then electric fields. Plan the various ground planes of the circuit board — analog, digital, and sometimes others — simultaneously with the layout. The analog ground plane (one or two) should be just below the outer layer(s) to minimize inductance of vias used for ground connections.
• Traces:
Keep traces in their respective analog or digital ground plane regions of the board if possible. Digital signals in analog areas may use an RC filter where crossing the analog-digital boundary — for example, in control of analog switches where longer edge transitions are acceptable. Consider that vias normally go all the way through the board. Once, I found a via with a digital signal that added noise to a MEMS gyro analog circuit, causing the product to miss the noise spec.
The old adages, “Don’t run noisy and noise-sensitive traces in parallel” and “Cross them only at right angles” will probably not even present themselves if layout was done properly.
• Shielding:
Sensitive sensors can often be placed in a conductive shield or enclosure, to shield from noise. If sensors are near or on the main analog circuit board, connect the shield to the circuit board’s analog ground at several points for lower inductance. Aluminum is usually used in a shield for ease of manufacture. Note that aluminum is normally anodized, so a good method is needed to break through the insulating layer. While aluminum conducts well and shields well for electric and electromagnetic fields, it’s less effective for magnetic, although induced eddy currents may help. If magnetic noise is a problem, see a relative permeability table for other materials because aluminum has about 1.000 — that of air.
• Impedances:
To reduce pick-up of radiated noise on the circuit board, keep resistances low. (This also reduces noise from resistors and noise from input noise currents of op amps.) Lower resistances increase power consumption and heat. Fortunately, the standard for supply voltages for precision analog has changed from ±15 V to ±5 V during the decade of the 2000s, reducing power by a factor of 9 (for the same resistor values). Signal range has only been reduced by a factor of about 2, from a standard of ±10 V to a few hundred mV from the power rails, depending on the op amp type and the output current. Using rail-rail range for the signal range can introduce errors into the signal.
Keeping heat and power efficiency in mind, other ways of reducing radiated noise pickup should also be used (such as shielding) thus enabling higher resistances and therefore less current, less heat, and lower wattage resistors. An exception is driving cables, especially if long and unshielded, because cables can act as antennas.
Transimpedance amplifiers have a high impedance at the op amp inverting input, and are critical for achieving low noise. The signal originates from a photo-sensor or other sensor that is a current source and is ideally infinite in impedance. A high value feedback resistor for the transimpedance amplifiers is needed for optimum S/N ratio because additional op amp gain stage(s) that are appended will result in higher S/N for the same over-all gain. The feedback resistor’s upper limit is set by the required signal bandwidth. The higher the parasitic capacitance of the sensor, and/or the lower the gain bandwidth (GBW) of the op amp per data sheet, the lower this feedback resistance needs to be, to achieve the same bandwidth. There is a non-linear equation that relates these parameters. A SPICE simulator should also be used to plot frequency response and noise spectrum or noise power spectral density (PSD) and to calculate total noise.
• Resistors:
For lower noise use metal film or thin film resistors — today, 0603 size is common — and avoid carbon composition and thick film. Resistor noise in Vrms is at least = sqrt(4*K*T*R*B). Carbon and thick film have additional noise. Accuracy is also much better for metal or thin film (.05% to 1%, not 5%) with better tempco (10 to 100 ppm/degree C, not 200).
Considering the price today is about 2 cents for 1% resistors in reel quantity (typically, 5000), it’s usually more economical to stock 1% resistors (for stockroom and production) as a default for analog circuits rather than use both 1% and 5%, considering ordering costs, stocking costs and assembly costs — fewer reels are cheaper. Buy tighter tolerance where needed.
• Power supplies:
Switching power supplies — all DC-DC and most AC-DC today — are very noisy and cannot directly power precision analog circuits. The noise can be greatly reduced (at the circuit board level) by using LC filtering in three frequency bands, each being low-pass filters in topology. This involves use of coils (up to, typically, 2 MHz), large ferrites (typically, 1 to 50 MHz) and small inductive chip beads (30 to 300+ MHz), along with ceramic capacitors. Look at the data sheets for impedance versus frequency for all inductors and capacitors, to calculate attenuation. These components only work below their self-resonant points. Be careful not to approach saturation current of any of these three types of inductors; otherwise their inductance would reduce. Place in parallel if need to reduce current.
Small chip inductive beads have both reactive and resistive components of impedance as a function of frequency. The resistive portion may eliminate the resonance of an LC efficiently, if it occurs, by providing damping. The capacitance of ceramic capacitors can be greatly reduced by a large DC voltage across it (as percentage of rated voltage). This phenomena becomes greater when using smaller packages for the same capacitance and voltage rating.
After the LC filters, use linear voltage regulators that have high noise rejection from input to output. Positive voltage regulators have much better rejection than negative, but some negative regulators types can increase noise rejection by 10 to 20 dB by placing a capacitor for added feedback, as described in their data sheets. It’s well worth using 0.1% resistors (and tempco of 10 or 25 ppm/degree C) if using resistors to set the voltage, since the cost is small compared to the regulator IC (typically, 1% to 1.5% accuracy). Doing worst-case spreadsheet analysis, tighter rail tolerances reduce worst-case power in op amps and the regulator. For example, using 1% rated resistor to set voltage, the voltage tolerance actually becomes 3% to 4% because there are 2 resistors with tempco. 4% of 5V is 0.2V which is 40% of a 0.5 V LDO (low drop-out regulator). Raising the preceding DC-DC voltage by 0.2V to compensate, adds 200 mW to heat generated in the LDO with 1 A current, and 8% more power consumed by op amps (for the same signal range).
If using a remote Power Distribution Unit or AC-DC power supplies to power the circuit board containing the precision analog circuits, the “DC” power cable could radiate noise into your signal cables, sensors and/or circuit board. The radiated noise is not only at the switching frequency, which typically is 50 kHz to 500 kHz, it also extends at harmonics of that, only very gradually diminishing with increased frequency. If this occurs, place a filter — inline if need be — at this power supply output. If inline, then design your own. Packaged AC power line filters are not effective because they need to pass AC efficiently. Alternatively, use twisted, shielded cable for the power cable. On the analog circuit board, place an isolating DC-DC converter for each rail voltage required, followed by LC filters and voltage regulators described above. In addition to noise, "DC" input to the circuit board will be degraded by power cable resistance and inductance, adding noise and fluctuating voltage.
• Signals and cabling between boards and subsystems
The grounds of different boards and subsystems are not going to be the same even if shorted to each other, excepting, for practical purposes, boards that are nearby, isolated, low current and not RF. The further removed the boards are from each other and/or the greater the current flow and/or the higher the frequency of signals, the more pronounced the difference. If this variation is great, it risks signal clipping and/or component failure. In such cases, prevent failure of signaling components by using protection diodes to limit ground variation between board grounds so that maximum voltage ratings of components, such as op amps, will never be exceeded. Typically, these are two or three diodes in series connecting the two board grounds, and repeat with diodes reversed, thus allowing a maximum of two or three diode drops for the ground difference. These diodes do not conduct during normal operation, but during a failure, power on/off transient, or static discharge if a subsystem became disconnected from ground.
• Analog signals between boards:
Use differential signaling, which has two conductors, to solve the problem that the grounds of the two boards or subsystems are different. If one of the conductors of the signal is connected to source ground or to a reference voltage, that is called unbalanced differential. If the two conductors are at equal voltage with opposite polarity, respect to source ground, that is balanced differential. The cost of balanced differential, versus unbalanced, is an extra op amp at the source. However, balanced differential has these advantages:
(1) eliminates common mode error of the receiving op amp, or at least greatly reduces it if the two grounds have a small voltage difference (the usual case),
(2) doubles S/N ratio where N is noise radiated into the cable,
(3) reduces crosstalk when signals are bundled in a cable, and
(4) enables the polarity of the signal to be easily reversed without causing saturation in the receiving circuit.
The last item is best achieved using the resistors for cable’s source impedance, which are in series with the driving op amp outputs. Place two extra (unmounted) resistors in the layout of the source board so the four resistors (two unmounted) also act as jumpers for polarity. Alternatively, an analogous scheme can be used at the receiver, placing two extra chip ferrites (for eliminating RFI from the cable using LC) so the four (two unmounted) act as jumpers. Swapping wires in the cable would also swap polarity, but entails more work and may be kludgy. During system integration, polarity errors are very common because when a number of engineers design circuit boards, motor placement, sensor connections, etc. on a project, they often differ in what is a positive direction for mechanical position, rotation, light intensity, and other physical phenomenon.
To achieve more protection for analog signals going between boards or subsystems, there are specialty isolating ICs for analog signals. However, these add additional noise, inaccuracies, and bandwidth limits (and costs).
If additional noise can be tolerated, an RC filter using a high resistance often suffices, along with diodes (which are often internal to ICs).
• Cabling for analog signals
For highest accuracy at low frequencies or short cables (< wavelength / 40), avoid current in the wires to prevent I*R drops by using, at the receiving end, either an inamp (instrumentation amplifier) — most of which work below 1 MHz — or two voltage follower op amps plus a differential amp stage — which also work at higher frequencies. This also eliminates inaccuracies of source and load resistors.
For higher frequencies and/or long cables, the driving and receiving impedance should match that of the cable. This means using source and load resistors, typically 50 ohms. The resistors reduce accuracy, but they can be 0.1% in 10 or 25 ppm/degree C. These cost about 5 cents and 30 cents, respectively, in reel quantity and 0603 package size. There is also I*R drop from current flow in the cable, but that’s small compared to the nominal 50% drop in signal amplitude for using these resistors. An alignment procedure using software can overcome constant (fixed) inaccuracies in most systems. However, this excludes temperature effects. To drive the cable, use an op amp that has high current output in the required bandwidth and signal range.
Keep these cables away from switching power supplies, solenoids, and other magnetic noise sources, which are much harder to shield from than electric field noises.
For differential signaling, use twisted-pair wires in most cases. Consider that one shield in common — which surrounds the cable — can introduce crosstalk between channels due to the capacitance of the shield to the wires. Twisted pair cable can be either cylindrical or flat. Cylindrical cable may have more crosstalk — unless pairs are individually shielded.
Short cables may use parallel flat cable. Closely spaced conductors (such as FFC with 0.5 mm pitch) have less noise pickup.
For flat cable, either twisted-pair or parallel (not twisted), use plenty of conductors as ground, ideally, between each twisted pair or signal pair. Connect both ends of these grounded conductors if one of the boards has complete isolation for its signals and power, and the cable is part of the “ground connection”. Isolation means there should be insignificant current flowing in the grounded conductors. Otherwise, ground at the source end in most cases. Providing grounded conductors this way reduces crosstalk and radiated noise pickup, and reduces impedance.
To obtain the highest S/N for one individual signal, use twisted-pair, individually shielded cable with braided shield.
• Digital signals between boards:
Unless the digital signal current into the receiver is almost zero, use a digital isolator to avoid ground loops that create noise in the analog section. (Some signal standards return most of the current, such as LVDS, so only the non-returned current needs consideration.) These isolators come in 3 basic types: capacitor coupled, transformer coupled, or magnetically coupled. The maximum frequency varies up to about 100 MHz. For the capacitor or transformer types, the DC and low frequency components of the signal is separated from the higher frequencies using filters, then modulated, transmitted across the isolation barrier, received, converted back to DC-low frequency, and then summed with the higher frequency components of the signal, thus reconstructing the original signal, all in a convenient IC package.
Optical types (opto-couplers) can be used for lower frequency signals.
• Ground planes of a circuit board:
• Connecting ground planes:
A common question is where and how the digital and analog ground planes should be connected on the circuit board, in a common situation where there are ADCs and a processor. The answer is to use a 0-ohm resistor to connect the planes, and select the footprint of that so that a chip (ferrite bead) inductor, selected before-hand, can be substituted if needed, to connect the planes. (Place empty pads for additional inductors in parallel, if needed, to keep below inductor saturation current.) If the system and board design was as described in this article, the 0-ohm would give the lowest noise at the ADC. If there is a ground loop, the inductor may reduce noise seen by software. The 0-ohm resistor also causes the netlist, generated by the schematic program, to be correct, rather than generate an error of shorting two nets (Analog GND and Digital GND).
The location of this 0-ohm connection is described below under ADCs.
Another consideration is that, assuming there are power planes and ground planes, the return currents in the ground plane want to travel directly over (or under) the path that is taken by the same current in the power plane, originating from the voltage sources. Therefore, the ground planes should extend over (or under) this current in the power plane(s), or else inductance will be added to the ground and power planes. To minimize capacitive coupling of digital noise into the analog plane, only extend power and ground planes over their respective analog or digital areas, not the entire board.
• ADCs:
The design is different depending on whether the ADC is internal to a processor IC or is an ADC IC only.
Locate ADCs — those not in a processor — near or at the analog-digital boundary, preferably near where they are connected by the shorting jumper connecting the ground planes. Connect the capacitors of the anti-aliasing filters to analog ground. Most ADCs have a pinout that separates analog and digital, so that analog components and traces can be placed over the analog ground plane and digital ones over the digital plane.
If an ADC is inside a processor, use a separate “mini-ground-plane” to connect the anti-aliasing filter capacitors, and then tie this plane to the digital ground plane at one point using a 0-ohm resistor, close to the processor. Locate resistors of the RC anti-aliasing filters at the boundary to minimize digital noise pickup on the analog signal, and connect the capacitor to the mini-ground-plane.
If an ADC is inside a processor and if digital and analog grounds are different, such as if using an inductor to separate the planes, digital noise can be added onto the analog signals, even with a large R (10 k ohm) in the anti-aliasing filter. This occurs if the op amp driving the analog signal has a low bandwidth, which is typical for precision op amps, or if the signal has a high source impedance. There are solutions: (1) connect the grounds together at or near the processor/ADC, or (2) use an ADC that is not in a processor, or (3) add an extra RC filter stage with C to analog ground, prior to the anti-aliasing filter (with its C to the mini-ground-plane, described above).
• Testing:
Measure the noise. For frequencies under 100 kHz, a Dynamic Signal Analyzer has highest accuracy for noise measurement. An FFT Analyzer is also good and is similar but has just one channel. For higher frequencies, use a Spectrum Analyzer. Measurement of noise by an oscilloscope, using the FFT function, will not have the frequency differentiation and dynamic range of the above instruments.
When designing the schematic, place an extra op amp on the board to amplify and observe the noise of the power rails, and possibly some signals, as well. Use an inverting summing op amp topology with AC-coupled inputs, but only connect one capacitor at a time, to select which input to measure. A gain of 10 to 30 should be fine to observe on the scope, but at least match the highest gain bandwidth (GBW) amplification stage in your system (e.g., use the same op amp and not higher gain) — or else the observed noise would be attenuated. Connect the output to a test point on the schematic.
Connect an oscilloscope to that test point. You should not observe any power supply switching noise on your rails if the three-band filter and voltage regulator was used, as described above. Analyzers have greater sensitivity so switching noise may be observed.
Oscilloscopes usually display some noise caused by the way the scope is grounded, sometimes called “artifact noise”, but this noise is not really present. This non-existent noise is seen by connecting the scope probe to the analog ground of the circuit board, at or near the scope ground-probe-clip. To this, add the noise pickup of the scope cable — seen by shorting the probe to probe-ground-clip and waving it around in the air — it should vary. When placing the scope probe on a signal on the board, the displayed noise will be the sum of these noises and the noise of the signal. Therefore, if the noise of the signal is less than the artifact noise, it cannot be measured using the oscilloscope.
Another common technique for noise measurement, where the signal measured is brought to an ADC, then a processor, is for software to compute the noise of the signal — with no actual signal present — and display the result in Vrms or other units.
Finally, place plenty of test points, including to the various grounds. It’s amazing how many engineers — even very experienced ones — forget to add test points to the schematic before going to FAB. This results in a lot of wasted time during the test and debug process: Looking for probing points on the physical board, clips falling off IC pins, and a lack of enough hands to hold the various probes if it can’t be clipped on. Soldering temporary wire test points can break or crack the component, such as a chip resistor or capacitor, just from the weight of the probe.
Write the test and calibration procedure — at least the first draft of it — before going to FAB. You will probably think of additional test points and jumpers needed on the board, needed for efficient debugging and production testing or for options, by writing that. You may even think of improvements to the design as you take a second look at it, by writing how to test and calibrate it.
Simple actions can be taken at the schematic level to make debug and production test both more efficient and with fewer errors. For example, don’t let the schematic or layout software decide how to number the test points. Number them in a logical manner, to save testing time. If you have 7 channels of analog signal that are similar in function, you can number the input test points 11 to 17, intermediate signals 21 to 27 and output 31 to 37, respectively, for channels 1 thru 7. All power rails need test points, and also DC-DC outputs. These can be numbered in the 50s. Place liberal analog grounds, maybe numbered in the 60s and 70s, digital ground in the 80s. The most common use of this scheme is connecting probe clips to a nearby ground — thus saving time when looking up test point numbers. Such simple actions can save many man-hours over the life of the product, and reduce errors such as putting the probe on the wrong test point.
——————————————————————————————————————————————————
Akos Szoboszlay holds a BS EE from San Jose State University. Born in Hungary, Akos has lived in Silicon Valley since 7th grade. Akos enjoys mountain biking, hiking, travel, and conversing in Magyar (Hungarian). To contact Akos (pronounced “Ah-kōsh”), go to the main Low Noise Design page, at the low noise design page .