Menu

The Ground That Isn't a Ground

The Ground That Isn't a Ground

Notes · Circuit boards and EMCNovember 19, 202624 min readSenova Systems Research

On the schematic every ground symbol is the same point. On the board they are joined by copper, and copper has resistance and inductance, so the ground at the sensor is not the ground at the ADC. The fix is a plane, a return under the trace, and one bridge between the noisy and the quiet.

The sensor readings jitter. The ADC is 100 mV off. The board fails EMC. And the schematic is correct.

That is the sentence that starts this article, because it is the sentence that starts the fault. The schematic is correct. The sensor is correct.

The ADC is correct. The code is correct. And the readings are wrong, because the ground is not a ground.

On the schematic, every ground symbol is the same point. On the board, they are joined by copper, and copper has resistance and inductance, so the ground at the sensor is not the ground at the ADC. The ground at the sensor is 100 mV above the ground at the ADC, and the ADC measures everything from its own ground, so its readings are 100 mV off.

This article is about the piece of copper. It is about the ground as a component, with resistance, with inductance, with a voltage drop, and about the three things that fix it: a plane, a return under the trace, and one bridge between the noisy and the quiet.

The simple idea first

Think of the ground as a road. On the schematic, the road is a line on paper, and every house on the road is at the same address, zero volts. On the board, the road is a strip of copper, and the strip has resistance, so the far end of the road is not at the same voltage as the near end.

The current that flows down the road drops voltage along it, the way a river drops height along its bed. The ground at the far end is higher than the ground at the near end, by the current times the resistance.

Now think of the road at high frequency. The strip of copper also has inductance, and inductance is a resistance that grows with frequency, the faster the current changes, the harder it is to push through. At 10 MHz, a thin strip of copper that is 0.1 Ω at DC is 3.1 Ω of reactance.

The ground at the far end is not just higher, it moves, up and down, with every switching edge, by a third of the signal's swing. That is ground bounce: noise on every signal the chip produces or reads, caused by the ground itself.

The fix is to make the road wider. A ground plane is a whole layer of copper, so wide and so close to everything that ground really does stay at nearly one voltage: milliohms for steady current, and about a nanohenry for fast current. (LearnEE) The plane is not a better road. It is the whole field, and the current takes the shortest path across it, which is the strip directly under its signal trace, so the loop is tiny, about 1 nH, and the ground barely moves.

The DC ground: the 100 mV that is not in the schematic

A ground plane is a whole layer, or a large poured area, of a circuit board's copper connected to ground. On a schematic every ground symbol is the same point. On a board they're joined by copper, and a ground plane makes that copper so wide and so close to everything that ground really does stay at nearly one voltage: milliohms for steady current, and about a nanohenry for fast current. (LearnEE)

A thin ground trace snaking back to the supply might be 0.1 Ω. A motor's 1 A through it raises the ground at a chip by 100 mV, and the chip measures everything from there, so its readings are 100 mV off. A plane, as wide as the board, is a few milliohms. (LearnEE)

Read that again, slowly, because it is the whole fault in two sentences. The trace is 0.1 Ω. The current is 1 A. The voltage is 100 mV. And the chip measures everything from its own ground, so the chip's readings are 100 mV off, and nothing in the schematic says so, because the schematic draws the ground as a point, and the point has no resistance.

The hub's sensor is an ADC, and the ADC measures the sensor's voltage against its own ground. The sensor's ground and the ADC's ground are joined by a trace, and the trace carries the load current, and the load current drops voltage along the trace, so the sensor's ground is above the ADC's ground by the current times the resistance. The ADC reads the sensor's voltage minus the ground shift, and the ground shift is the error.

How big is the error? The hub's sensor ADC is 12-bit on a 3.3 V rail (the article's own worked example, the ADC resolution is a labelled assumption). One least significant bit, one count, is 3.3 V divided by 4,096, which is 0.81 mV.

A 100 mV ground shift is 124 counts. The sensor's full scale is 4,096 counts, so 124 counts is 3.03% of full scale. The ADC can resolve 0.81 mV, but the ground moves by 100 mV, 124 times the resolution, so the resolution is wasted.

The ADC is a 12-bit instrument measuring a ground that moves by 124 counts.

Article #38 said the same thing about a different instrument: the Solinst Levelogger 6 M100 has ±5 cm accuracy and 0.6 mm resolution, the resolution is 83 times finer than the accuracy. The resolution is not the accuracy. The ADC's 0.81 mV resolution is not the measurement's accuracy, because the ground moves by 100 mV, and the ground is part of the measurement.

The plane fixes the DC error. A 1-ounce copper plane is approximately 500 µΩ per square, so a 1 A change can produce about 500 µV per square of resistive drop. (RottenWifi, from Article #44) The sensor's ground path on the plane is two squares, 1 mΩ, so at 1 A the drop is 1 mV, not 100 mV.

One millivolt is 1.24 counts, not 124. The plane reduces the DC ground error by 100 times, from 124 counts to about 1 count, because the plane is wide and the trace is thin, and resistance is inversely proportional to width.

The AC ground: ground bounce

The same trace has about 50 nH. To a fast current that's a reactance that grows with frequency: X = 2π f L = 2π × 10 MHz × 50 nH ≈ 3.1 Ω. A chip driving a 10 Ω load at 10 MHz through that ground moves its own ground voltage by about 30% of its swing. That's ground bounce: noise on every signal the chip produces or reads. (LearnEE)

At 10 MHz the thin trace moves the ground voltage by 30% of the clock's swing; the plane by 0.6%. (LearnEE)

The arithmetic behind LearnEE's "about 30%" is the article's own: the trace's reactance at 10 MHz is 2π × 10 MHz × 50 nH = 3.14 Ω, and 3.14 Ω against a 10 Ω load is 31.4% of the load, so the ground moves by about a third of the swing. The plane's inductance is about 1 nH, so its reactance at 10 MHz is 2π × 10 MHz × 1 nH = 0.063 Ω, and 0.063 Ω against a 10 Ω load is 0.63%, LearnEE's 0.6%.

The trace moves the ground 50 times more than the plane, because 50 nH is 50 times 1 nH, and reactance is proportional to inductance.

And the hub's clock is not 10 MHz. The hub's STM32H7 runs at 48 MHz (from the cluster). At 48 MHz, the trace's 50 nH is 2π × 48 MHz × 50 nH = 15.08 Ω, 150.8% of a 10 Ω load, which means the ground would move more than the signal itself.

The plane's 1 nH at 48 MHz is 0.302 Ω, 3.02% of the load. The trace is not just worse than the plane at the hub's clock. The trace's ground impedance is larger than the load it is supposed to serve.

This is why the ground grid exists, and why it stops at 10 MHz. A study compared the voltage difference between the ground pins of the ICs of a two-layer board that uses a ground grid to that of a similar design with a single-point ground. The study shows that the ground grid can reduce the voltage difference between the ground pins of the ICs from 1000 mV to 100 mV (this was the best improvement between the various combinations of the ICs).

According to this study, the board radiation was reduced by about 7 dB when the ground grid technique was applied. (All About Circuits, from Henry W. Ott, Electromagnetic Compatibility Engineering §10.5.3) The grid reduces the ground difference by 10 times, from 1,000 mV to 100 mV, 90% less, and the board's radiation by about 7 dB, which is 10^(7/10) = 5.01 times less radiated power (the article's own arithmetic).

To achieve a significant reduction in the ground circuit inductance, a ground grid spacing of 0.5 in. or less should be used. (Ott) Half an inch is 12.7 mm. The ground grid technique is compatible with circuits operating up to 10 MHz; above this frequency a ground plane should be used.

(Ott) The grid is the fix for a two-layer board at 10 MHz and below. Above 10 MHz, the fix is the plane, and the hub's clock is 48 MHz, so the hub's board is four layers, with the ground layer right under the parts.

The return current: the path of least inductance

At high frequencies, a return current doesn't take the shortest path back. It takes the path of least inductance, which is the smallest loop: in a plane, the strip directly under its signal trace. That loop is tiny, about 1 nH, so the ground barely moves, and the trace and its return radiate far less. (LearnEE)

This is the sentence that makes the plane work, and it is the opposite of what the intuition says. The intuition says the current takes the shortest path, the straight line back to the supply. The physics says the current takes the path of least inductance, the smallest loop, and the smallest loop is the strip directly under the signal trace, because the signal trace and its return are close together, and the loop area between them is small, and inductance is proportional to loop area.

Article #19 said the same thing about radiation: the differential-mode loop radiates as E = 1.32·A·I·F²/r, where A is the loop area, so the smaller the loop, the less it radiates. The return under the trace makes the loop small, so the trace and its return radiate far less. And Article #44 said the same thing about the SMPS: the hot loop is the loop the switching currents travel in, and the goal is to minimize the size of that loop, because the loop is the antenna.

The ground plane is the thing that makes the loop small, because the return can run directly under the signal, instead of snaking back through a thin trace.

Three rules follow, and they are the whole layout discipline:

  • Connect the plane at both ends. A plane connected only at the chip carries no current; the return current needs it connected at the supply too. (LearnEE)
  • Stitch with vias. On two-layer boards, vias tie top-side ground pours to the bottom plane so returns can switch layers. (LearnEE)
  • Decouple into it. A decoupling capacitor's short path to the plane is what lets it feed a chip's fast current. (LearnEE)

And one more, from Article #44: the three grounds, the input capacitor's ground, the device's ground, and the output capacitor's ground, should be as close together as possible, connected with a solid ground plane on the same layer as the components, so the violent switching currents avoid the internal PCB ground plane entirely. The decoupling capacitor's short path to the plane is what lets it feed the chip's fast current, and the short path is short because the capacitor sits on the plane, not at the end of a trace.

Single-point or multi-point: the frequency decides

In low-frequency circuits where the signal frequency is less than 1MHz and the inductance between routing and components can be ignored, the voltage drop on the ground line circuit resistance has a significant impact on the circuit, so a single-point grounding method should be adopted. When the signal frequency exceeds 10MHz, the influence of ground line inductance is significant, so it is advisable to use multi-point grounding closer to the signal source.

When the signal frequency is between 1~10MHz, if a single-point grounding method is used, the length of the ground line should not exceed 1/20 of the wavelength; otherwise, multi-point grounding should be adopted. (PCBbuy)

The rule is a frequency rule, not a preference. Below 1 MHz, the resistance matters and the inductance does not, so the ground is a single point, one star, one connection, no loops, because a loop is a path for circulating current, and circulating current is a ground difference. Above 10 MHz, the inductance matters and the resistance does not, so the ground is multi-point, many connections, close to the source, because every millimetre of trace is nanohenries, and nanohenries at 10 MHz are ohms.

Between 1 MHz and 10 MHz, the rule is a length rule: the ground line should not exceed 1/20 of the wavelength. The wavelength at 1 MHz is 300 m, so 1/20 is 15 m, longer than any board, so a single point is fine. The wavelength at 10 MHz is 30 m, so 1/20 is 1.5 m, still longer than any board, so a single point is fine on a board, and the rule bites on cables, not on boards.

The 1/20 rule is the article's own arithmetic, from the speed of light: wavelength = 3×10⁸ / frequency, so 1/20 of the wavelength at 10 MHz is 1.5 m, and at 1 MHz it is 15 m. The board is 100 mm long. The ground line on the board is shorter than 1/20 of the wavelength at any frequency the board uses, so the single-point rule is not the board's problem.

The board's problem is the inductance, and the inductance is fixed by the plane.

The ground line should be as wide as possible, generally greater than 3mm. (PCBbuy) The ground wire should be as thick as possible so that it can pass three times the allowable current of the printed circuit board. (PCBbuy) Three millimetres is the minimum width for a ground trace, and three times the allowable current is the minimum capacity, because the ground is not a signal, it is the return for every signal, and every signal's current flows through it.

Split grounds: the noisy and the quiet

When a board has noisy and quiet sections, split grounds are sometimes used, joined at exactly one bridge, under the point where analog and digital meet. (LearnEE)

The split is the answer to the question the plane cannot answer: what if the ground has to be both noisy and quiet? The digital section is noisy, the processor switches, the clock edges are fast, the ground bounces by 3% of the swing at 48 MHz. The analog section is quiet, the sensor's ADC measures microvolts, and the analog circuit portion may malfunction with noise at the microvolt level.

(PCBbuy) The digital circuits have strong anti-interference capabilities, with a noise tolerance of 0.4–0.6V for TTL circuits and 0.3–0.45 times the power supply voltage for CMOS digital circuits, while the analog circuit portion may malfunction with noise at the microvolt level. (PCBbuy) The digital section tolerates 400–600 mV of noise. The analog section fails at microvolts.

The same ground cannot serve both, unless the grounds are split, and joined at exactly one bridge, under the point where analog and digital meet.

The bridge is the single point where the two grounds meet, and it is under the point where the analog and digital signals cross, so the return currents do not cross each other. The digital return current flows through the digital ground, back to the supply, without passing under the analog section. The analog return current flows through the analog ground, back to the supply, without passing under the digital section.

The two grounds meet at one point, and only one point, so there is no loop, and no circulating current, and no ground difference between the two sections, except the difference across the bridge, which is the point where the signals cross, and where the signals cross, the ground difference is the signal's own reference.

The mistake is to split the grounds and join them in two places, or to split them and not join them at all. Two joins make a loop, and a loop is an antenna, Article #19's differential-mode loop, E = 1.32·A·I·F²/r, where A is the loop area. No join makes two grounds, and two grounds are two different voltages, and the ADC measures the difference between them. One bridge, under the point where analog and digital meet, is the whole rule.

The shield: the enclosure is the other half of the ground

The ground plane is the shield inside the board. The enclosure is the shield outside it. And the hub's enclosure is a hybrid, Article #47's answer: an aluminium base and a polycarbonate lid. The aluminium base is the shield. The polycarbonate lid is not.

Metal enclosures attenuate wireless signals. Plastic enclosures provide negligible EMI shielding unless coated with a continuous, grounded conductive layer. (Fabcon, from Article #47) The aluminium base attenuates the wireless signals, it is a shield, because aluminium is a conductor, and a conductor is a ground plane in three dimensions.

The polycarbonate lid provides negligible EMI shielding, because polycarbonate is an insulator, and an insulator is not a ground. The hub's lid is polycarbonate because the BLE radio has to transmit through it, the radio is at 2.4 GHz, and the aluminium base would block it. The hub's base is aluminium because the board has to be shielded, and the base is the shield, and the base is grounded to the board's ground plane, so the base and the plane are one ground.

The 4-layer board produces about 15 dB less radiation than a 2-layer board. (Article #19) Fifteen decibels is 10^(15/10) = 31.6 times less radiated power (the article's own arithmetic). The hub's board is four layers, signals, ground, power, signals, with the ground layer right under the parts (LearnEE), so the board radiates 31.6 times less than a two-layer board would, and the aluminium base shields what is left.

The ground plane and the enclosure are the same idea at two scales: the plane is the shield inside the board, the enclosure is the shield outside it, and both work because they are conductors, and conductors are grounds.

The shield is only as good as its seams, and the hub's seam is the lid-to-base joint, held by 8 × M4 screws (from Article #49), opened 25 times over a 10-year life (the labelled assumption from Article #49). Every time the lid opens, the joint gets weaker, Article #49's whole argument, and the shield's seam is the joint. The screws are the shield's seam, and the inserts are the shield's strength, and the ground plane is the shield's floor.

The enclosure, the fasteners, and the ground are one system, and the system is only as good as its weakest joint.

The ground fault on the hub board

The hub board is 100 × 60 mm, 4-layer, 120 nets, 114 test points. (Article #32) The hub's STM32H7 runs at 48 MHz. The sensor's ADC is 12-bit on a 3.3 V rail (the article's own worked example, the ADC resolution is a labelled assumption).

The hub's peak draw is 100 mA (the article's own worked example, a labelled assumption). The ground trace from the sensor to the ADC is a thin trace, 0.1 Ω, about 50 nH (LearnEE's numbers, applied to the hub board). The question is: how wrong is the sensor's reading?

Step 1, the DC error. The thin trace is 0.1 Ω. The hub's peak draw is 100 mA. The ground shift is 0.1 Ω × 0.1 A = 10 mV. One LSB is 3.3 V / 4,096 = 0.81 mV. The ground shift is 10 mV / 0.81 mV = 12.41 counts, 0.303% of full scale. The ADC can resolve 0.81 mV, but the ground moves by 10 mV, 12 times the resolution, so the sensor's reading is 12 counts off, and the 12 counts are not noise, they are the ground. Compare LearnEE's general case: a motor's 1 A through the same 0.1 Ω trace raises the ground by 100 mV, 124 counts, 3.03% of full scale. The hub's 100 mA is a tenth of the motor's 1 A, so the hub's error is a tenth of the motor's, 12 counts instead of 124, but it is the same fault, and the same trace.

Step 2, the AC error. At the hub's 48 MHz clock, the trace's 50 nH is 2π × 48 MHz × 50 nH = 15.08 Ω, 150.8% of a 10 Ω load. The ground would move more than the signal itself. The plane's 1 nH at 48 MHz is 0.302 Ω, 3.02% of the load. At LearnEE's 10 MHz, the trace is 3.14 Ω, 31.4% of the load, LearnEE's "about 30%", and the plane is 0.063 Ω, 0.63%, LearnEE's "0.6%". The trace moves the ground 50 times more than the plane at any frequency, because 50 nH is 50 times 1 nH. At the hub's 48 MHz, the trace's ground impedance is larger than the load, so the ground is not a reference, it is a signal, and the signal is the ground's noise.

Step 3, the plane's DC error. The sensor's ground path on the plane is two squares of 1-ounce copper. One square is 500 µΩ. Two squares are 1 mΩ. At the hub's 100 mA, the drop is 1 mΩ × 0.1 A = 0.1 mV, 0.124 counts. The thin trace's error was 12.41 counts. The plane's error is 0.124 counts. The plane reduces the DC ground error by 100 times, from 12 counts to about a tenth of a count, because the plane is wide and the trace is thin. The ADC's 12-bit resolution is finally usable, the ground moves by a tenth of a count, not by 12 counts.

Step 4, the grid, if the board were two layers. The ground grid reduces the voltage difference between the IC ground pins from 1,000 mV to 100 mV, a 10-times reduction, 90% less, and the board's radiation by about 7 dB, which is 5.01 times less radiated power. (Ott) The grid spacing is 0.5 in. or less, 12.7 mm. The grid works up to 10 MHz. The hub's clock is 48 MHz, so the grid is not the hub's fix, the plane is. The grid is the fix for the board that cannot afford four layers, and the grid's 1,000 mV to 100 mV is the measure of what a two-layer board can do, and what it cannot: the grid's 100 mV is still 124 counts on a 12-bit ADC.

Step 5, the 4-layer board. The hub's board is four layers, and a 4-layer board produces about 15 dB less radiation than a 2-layer board (Article #19), 31.6 times less radiated power. The ground layer sits right under the parts, so the return current runs directly under its signal trace, and the loop is about 1 nH, and the ground barely moves. The 4-layer board is not a luxury. It is the fix for a 48 MHz clock, because the grid stops at 10 MHz, and the hub's clock is 48 MHz.

Step 6, the fix, and the cost of finding it late. The fix is four things: the plane (the ground layer right under the parts); the return under the trace (the loop is 1 nH, not 50 nH); the split grounds joined at one bridge, under the point where analog and digital meet; and the decoupling capacitor's short path to the plane. The ADC's error falls from 12 counts to about a tenth of a count. The cost of finding the fault late is Article #40's cost escalation: 1× at design, 5× at prototype, 20× at pilot, 50× at production, 100×+ in the field, and Article #40's field failure, 4 of 10 hubs failed on days 19–23, 40% of the fleet, because the ground moved and the sensor's readings were wrong, and the readings were wrong because the ground was a trace, not a plane. The ground fault is a design fault, and the design fault is found in the field, and the field is 100 times more expensive than the design.

The decision test

Five questions. If the answer to any one of them is "we don't know", the ground is the fault, and the assessment is the next step.

  1. What is the ground impedance between the sensor and the ADC, at DC, and at the clock frequency? If the answer is "the schematic says ground", the ground has not been measured. The DC impedance is the trace's resistance, 0.1 Ω for a thin trace, 1 mΩ for two squares of plane. The AC impedance is the trace's inductance, 50 nH for a thin trace, 1 nH for a plane, times 2πf. At the hub's 48 MHz, the trace is 15.08 Ω and the plane is 0.302 Ω. The question is not whether the ground has impedance. It is how much, and at which frequency.
  2. Where does the return current flow? If the answer is "back to the supply", the return has not been drawn. The return flows through the path of least inductance, the strip directly under its signal trace, in a plane, or it snakes back through a thin trace, and the loop is 50 nH, and the ground moves. The return is drawn on the layout, not on the schematic, and the layout is where the fault lives.
  3. Are the analog and digital grounds split, and joined at exactly one bridge? If the answer is "we have one ground", the digital noise is in the analog section. The digital section tolerates 400–600 mV of noise. The analog section fails at microvolts. The bridge is under the point where analog and digital meet, and it is the only join. Two joins make a loop. No join makes two grounds.
  4. What is the ADC's error, in counts, from the ground shift? If the answer is "the ADC is 12-bit, so the error is 1 count", the ground has not been counted. The ground shift is the current times the resistance, 10 mV at the hub's 100 mA through a 0.1 Ω trace, and 10 mV is 12.41 counts on a 12-bit, 3.3 V ADC. The ADC's resolution is 0.81 mV. The ground's error is 12 counts. The resolution is not the accuracy.
  5. What does the ground fault cost if it is found in the field? If the answer is "a respin", the cost has not been counted. Article #40's cost escalation is 1× at design, 5× at prototype, 20× at pilot, 50× at production, 100×+ in the field, and Article #40's field failure was 4 of 10 hubs, 40% of the fleet, on days 19–23. The ground fault is found in the field, and the field is 100 times more expensive than the design.

Objections

"Our schematic is correct, the ground is ground." The schematic is correct, and the ground is not ground. On the schematic, every ground symbol is the same point. On the board, they are joined by copper, and copper has resistance and inductance. A thin ground trace snaking back to the supply might be 0.1 Ω, and a motor's 1 A through it raises the ground at a chip by 100 mV, and the chip measures everything from there, so its readings are 100 mV off. (LearnEE) The schematic is a drawing of the circuit. The board is the circuit. The ground fault lives on the board, not on the schematic, and the schematic cannot see it, because the schematic draws the ground as a point, and the point has no resistance.

"We can't afford four layers, the board is two layers and it works." The board works on the bench, at 25 °C, with no load, and the ground grid is the fix for a two-layer board at 10 MHz and below, the grid reduces the voltage difference between the IC ground pins from 1,000 mV to 100 mV, and the board's radiation by about 7 dB. (Ott) But the hub's clock is 48 MHz, and the grid is compatible with circuits operating up to 10 MHz; above this frequency a ground plane should be used. (Ott) The two-layer board works until the clock is faster than 10 MHz, or the load is larger than the grid can carry, or the field is hotter than the bench, and then the ground moves, and the sensor's readings are wrong, and the field failure is 40% of the fleet. The four-layer board is not a luxury. It is the fix for a 48 MHz clock.

"We split the grounds, the analog and digital are separate." The split is half the rule. The other half is the bridge, joined at exactly one bridge, under the point where analog and digital meet. (LearnEE) Two joins make a loop, and a loop is an antenna, Article #19's differential-mode loop, E = 1.32·A·I·F²/r, where A is the loop area. No join makes two grounds, and two grounds are two different voltages, and the ADC measures the difference between them. The split without the bridge is not a fix. It is two grounds, and the ADC is caught between them.

"The EMC lab will find it, that's what the lab is for." The lab will find it, and the lab will charge for it. About half of products fail EMC testing on the first attempt (Intertek Labs, from the batch-6 pipeline), and each retest cycle adds $5,000 to $30,000 and 2 to 6 weeks. (MarkReady) A pre-compliance scan catches the ground fault at $300–$800 per scan instead of $3,000–$8,000 per failed lab booking. (Ideambox) The lab is the test, not the fix. The fix is the plane, and the plane is drawn on the layout, and the layout is drawn before the lab booking, and the ground fault found in the lab is a respin, and the respin is a new board, and the new board is a new lab booking.

"The shield is the enclosure, the board doesn't need one." The enclosure is the shield outside the board, and the board needs a shield inside it, the ground plane. Metal enclosures attenuate wireless signals. Plastic enclosures provide negligible EMI shielding unless coated with a continuous, grounded conductive layer. (Fabcon, from Article #47) The hub's base is aluminium, so it shields. The hub's lid is polycarbonate, so it does not, and the lid is polycarbonate because the BLE radio has to transmit through it. The enclosure shields what the board radiates, and the board radiates 31.6 times less because it is four layers. The enclosure and the plane are the same idea at two scales, and neither replaces the other.

Sources

  1. LearnEE, "What is a ground plane? Why PCBs use one"
    What we used

    "A ground plane is a whole layer, or a large poured area, of a circuit board's copper connected to ground. On a schematic every ground symbol is the same point. On a board they're joined by copper, and a ground plane makes that copper so wide and so close to everything that ground really does stay at nearly one voltage: milliohms for steady current, and about a nanohenry for fast current." "A thin ground trace snaking back to the supply might be 0.1 Ω. A motor's 1 A through it raises the ground at a chip by 100 mV, and the chip measures everything from there, so its readings are 100 mV off. A plane, as wide as the board, is a few milliohms." "The same trace has about 50 nH. To a fast current that's a reactance that grows with frequency: X = 2π f L = 2π × 10 MHz × 50 nH ≈ 3.1 Ω. A chip driving a 10 Ω load at 10 MHz through that ground moves its own ground voltage by about 30% of its swing. That's ground bounce: noise on every signal the chip produces or reads." "At 10 MHz the thin trace moves the ground voltage by 30% of the clock's swing; the plane by 0.6%." "At high frequencies, a return current doesn't take the shortest path back. It takes the path of least inductance, which is the smallest loop: in a plane, the strip directly under its signal trace. That loop is tiny, about 1 nH, so the ground barely moves, and the trace and its return radiate far less." "Connect it at both ends. A plane connected only at the chip carries no current; the return current needs it connected at the supply too." "Stitch with vias. On two-layer boards, vias tie top-side ground pours to the bottom plane so returns can switch layers." "Decouple into it. A decoupling capacitor's short path to the plane is what lets it feed a chip's fast current." "When a board has noisy and quiet sections, split grounds are sometimes used, joined at exactly one bridge, under the point where analog and digital meet."

  2. PCBbuy, "Ground Line Design Considerations in PCB Design"
    What we used

    "In low-frequency circuits where the signal frequency is less than 1MHz and the inductance between routing and components can be ignored, the voltage drop on the ground line circuit resistance has a significant impact on the circuit, so a single-point grounding method should be adopted. When the signal frequency exceeds 10MHz, the influence of ground line inductance is significant, so it is advisable to use multi-point grounding closer to the signal source. When the signal frequency is between 1~10MHz, if a single-point grounding method is used, the length of the ground line should not exceed 1/20 of the wavelength; otherwise, multi-point grounding should be adopted." "Digital circuits have strong anti-interference capabilities, with a noise tolerance of 0.40.6V for TTL circuits and 0.30.45 times the power supply voltage for CMOS digital circuits, while the analog circuit portion may malfunction with noise at the microvolt level." "The ground line should be as wide as possible, generally greater than 3mm." "The ground wire should be as thick as possible so that it can pass three times the allowable current of the printed circuit board."

  3. All About Circuits, "Mitigating Noise in Double-sided PCBs: Multipoint Grounding (Gridded Ground) in Two-layer PCB Design"
    What we used

    (from Henry W. Ott, *Electromagnetic Compatibility Engineering* §10.5.3), "A study compared the voltage difference between the ground pins of the ICs of a two-layer board that uses a ground grid to that of a similar design with a single-point ground. The study shows that the ground grid can reduce the voltage difference between the ground pins of the ICs from 1000 mV to 100 mV (this was the best improvement between the various combinations of the ICs). According to this study, the board radiation was reduced by about 7 dB when the ground grid technique was applied." "To achieve a significant reduction in the ground circuit inductance, a ground grid spacing of 0.5 in. or less should be used." "The ground grid technique is compatible with circuits operating up to 10 MHz; above this frequency a ground plane should be used."

  4. Demo arithmetic verified in Python before writing: DC ground shift, thin trace 0.1 Ω × 1 A = 100 mV (LearnEE); 12-bit ADC on a 3.3 V rail → 1 LSB = 3.3 / 4,096 = 0.81 mV; 100 mV / 0.81 mV = 124.1 counts = 3.03% of full scale. Hub's own case, peak draw 100 mA (labelled assumption) × 0.1 Ω = 10 mV = 12.41 counts = 0.303% of full scale. Plane, 2 squares × 500 µΩ = 1 mΩ; at 1 A → 1 mV = 1.24 counts (100× less than the thin trace); at 100 mA → 0.1 mV = 0.124 counts (100× less). AC ground bounce, trace 50 nH at 10 MHz: X = 2π × 10 MHz × 50 nH = 3.14 Ω → 31.4% of a 10 Ω load (LearnEE's "about 30%"
    What we used

    ); plane 1 nH at 10 MHz: X = 0.063 Ω → 0.63% (LearnEE's "0.6%"); trace moves the ground 50× more than the plane. At the hub's 48 MHz clock: trace 50 nH → 15.08 Ω = 150.8% of a 10 Ω load; plane 1 nH → 0.302 Ω = 3.02%. Ground grid, 1,000 mV → 100 mV = 10× reduction, 90% less; 7 dB = 10^(7/10) = 5.01× less radiated power. 4-layer vs 2-layer, 15 dB = 10^(15/10) = 31.6× less radiated power. 1/20 wavelength, at 1 MHz: 300 m / 20 = 15 m; at 10 MHz: 30 m / 20 = 1.5 m. Grid spacing 0.5 in. = 12.7 mm

More in precision measurement

EMC pre-compliance testingAbout half of products fail EMC testing on the first attempt, and each failure costs $5,000–$30,000 and 2–6 weeks. A pre-compliance scan on your own bench finds the failure before the lab does, for $300–$800 a scan.30 minThe Rail That Doesn't RingYour schematic is correct and your board still fails EMC. The problem is in the layout: the hot loop, the switch node, and the ground that is not one ground. Here is the buck converter's noise, measured in numbers.20 minThe Rail That SagsThe processor wakes up, the rail droops, the board resets, the clock jitters, the radio loses packets, and the schematic is correct, because the problem is the impedance of the power delivery network, not the regulator.22 minSensor Calibration and DriftA gyroscope measures rate, not angle, the angle is the integral of the rate, so any bias is integrated too: a bias of 0.1°/s adds 6° of error every minute, 360° in an hour. Bias is the offset the sensor reads at rest; it shifts with temperature and time, and random noise adds its own slow wander (a random walk whose standard deviation grows with the square root of time). The fixes are simple and in order: measure the bias at startup while the device is still and subtract it; recalibrate continuously when the device rests; fuse with an accelerometer (complementary filter, α ≈ 0.98), the accelerometer gives tilt with no drift when still, the gyro gives smooth fast motion. Yaw has no gravity reference, so it needs a magnetometer. Even a well-calibrated gyroscope cannot be trusted for hours without a second sensor to check against.24 minThe Spec Sheet That Lies About AccuracyA sensor can be precise without being accurate, a tight cluster of readings, far from the true value. Accuracy is closeness to the true value; precision is how tightly repeated readings group; resolution is the smallest detectable change. The Levelogger 6 M100's resolution is 0.6 mm on a 100 m full scale, 83 times finer than its ±5 cm accuracy. Averaging fixes random noise; it cannot fix systematic error. That needs calibration against a known reference.22 min