Taller Than the Hum

1The bill arrives

Every move of the last three weeks had the same fine print. We shortened routes, deepened pipelines, and above all we turned the master dial down — because the voltage knob is squared, and nothing else pays like it. But look at what the dial actually sets: the height of a full bucket. The gap between "certainly a 1" and "certainly a 0" is made of volts, and every volt we shaved off the supply was shaved off that gap. We were not just buying joules. We were quietly selling margin — the safety gap — and tonight the buyer shows up to collect.

The buyer is noise, and the first thing to understand about it is that noise is not a malfunction. Nothing is broken. Nothing needs fixing. Noise is the ordinary background hum of physics itself: atoms jiggling because they are warm, neighbors whispering because wires sit close together, the supply wobbling because a million taps opened on the same tick. It was all there in week 2, when the hum was called k_BT and it was kicking a ball around a double well. It never left. It has been humming under every bench in this course, politely ignored, because our margins stood so far above it that it could not touch anything.

So the question of the week is not "how do we silence the hum" — you cannot; it is funded by temperature itself. The question is an accounting question, the kind this course likes: are the margins we kept still taller than the hum? By how much? And what exactly happens as the gap closes? The answer has a punchline you will want to frame: the gap does not close gracefully. It closes like a cliff.

Here is the route. Section 2 meets the noise, face by face. Section 3 is the heart of the night: the margin as an insurance policy, and the cleaning trick that makes digital computing possible at all. Section 4 brings week 2's barrier model home and answers the dial question with numbers. Section 5 explains the shaking neighbor. Section 6 counts the errors — and shows what engineers do when physics wins anyway.

2The rogues' gallery

"Noise" sounds like one thing. It is a small gang, and each member has a face, a motive, and a section of this course where you already met them. Three matter tonight.

The jiggle. Week 2's old adversary, now in plumbing terms. Every atom in the chip is warm, which means it is moving, and the moving atoms slosh the water in every bucket — the voltage on every node trembles, all the time, with no cause beyond temperature. Notes 03 priced the stake: about 26 meV of mischief per degree of freedom at room temperature, which on our standard one-femtofarad bucket works out to a tremble of about 2 millivolts. Tiny — and it never, ever stops. You can cool the chip to shrink it. You cannot turn it off.

The neighbor's whisper. Notes 09 taught you that a wire is a brigade of buckets — capacitance smeared along its whole length. Here is the part we postponed: capacitance does not only form between a wire and the ground beneath it. It forms between a wire and whatever sits next to it — including the neighboring wire. Two signal wires running side by side are an accidental capacitor, and a capacitor is a handshake: when one plate moves, it drags the other. Your neighbor flips, and your quiet wire twitches without anyone touching it. Section 5 measures the twitch.

The tank's wobble. One brief face, so you recognize it in the wild. The supply is a tank, and tanks are not infinitely stiff: when a million taps open on the same clock tick — which is exactly what a clock tick is for — the sudden withdrawal dips the tank's level for a moment, and every bucket filling at that moment fills a little short. Engineers call the dip droop and its mirror image on the ground side ground bounce. The fixes are plumbing (fat supply wires, little reservoir capacitors sprinkled everywhere), and we will leave it at that: just know that even the reference levels themselves wobble.

Stack the three and you get the total sloshing on a node — the jiggle underneath, whispers and wobble on top. On a quiet, well-laid-out chip the stack is a few millivolts. On a crowded bus next to a flipping neighbor on a droopy tick, it can be tens of millivolts. Hold that range; the benches use it.

3The insurance policy

Now the heart of the night. Digital design's answer to the hum is not a gadget — it is a contract, and you have seen its skeleton before: Notes 06 drew two lines on the inverter's cliff and called the zone between them forbidden. Tonight we read that drawing the way its authors meant it: as an insurance policy, written in volts, between every pair of gates on the chip.

The contract has two sides. The sender promises about its outputs: "when I say 1, I will put out at least this much voltage; when I say 0, at most this much." For our CMOS gates the sender's promises are luxurious — Notes 06 showed the inverter drives all the way to the rails, a full 1.0 V for a 1 and a true 0 V for a 0. The receiver promises about its inputs: "anything I see above my high threshold, I will read as a 1, guaranteed; anything below my low threshold, a 0, guaranteed." On the balanced inverter those thresholds sit at 0.56 V and 0.44 V. Between them is the forbidden zone: a no-man's-land where the receiver promises nothing at all.

And the gap between the promises — between what the sender guarantees to deliver and what the receiver demands to see — is the noise margin. It is not leftover slack. It is slack bought deliberately: a budget of volts that noise is allowed to steal from the signal, in transit, with no consequence whatsoever. The sender launches a 1 at 1.0 V; the hum can drag it down by anything up to 440 mV, and the receiver still reads it perfectly. That is insurance. You pay for it up front (in swing you could have shaved, hence in energy), and in exchange, ordinary bad luck is free.

Restoration — the trick that makes digital possible

Now the punchline, and it is the single biggest idea tonight. Look at what a CMOS gate does to a slightly dirty signal. A 0 arrives scuffed up — say the whispers shoved it to 0.38 V. That is ugly, but it is still below the 0.44 V threshold, so the receiver reads it as a 0 with certainty — and then, because its output is driven hard to the rails (Notes 06's cliff), it puts out a factory-fresh, full-strength answer: a crisp 1.0 V. The 380 mV of dirt is simply gone. Not passed along. Not averaged in. Erased. Every logic gate on the chip is secretly also a cleaning station, and engineers call this restoration: the signal is restored to full health at every single doorstep it passes.

Here is why you should believe this is the trick that makes digital beat analog at scale — no theorem needed, just two chains. Build a chain of a hundred analog stages, each adding its own 10 mV of hum to whatever it receives: the dirt has nowhere to go, so it piles up, stage after stage, and after a hundred stages you have roughly a volt of accumulated noise riding a one-volt signal. The message has drowned in the hum. Now build a hundred digital gates: each one receives dirt, checks it against the contract, and outputs clean. The noise never accumulates — the only dirt on the signal at any point is what it picked up since the last doorstep, one hop's worth, forever. A digital signal can cross a billion gates and arrive as pristine as it left, because it was never asked to survive a billion gates. It was asked to survive one hop, a billion times.

One honest caveat, because it sets up everything after this line: restoration is perfect only while the dirt stays inside the margin. A gate that reads 0.38 V cleans it flawlessly. A gate that reads 0.47 V — inside the forbidden zone — promises nothing, and a gate that reads 0.60 V when a 0 was sent doesn't clean the error, it launders it: out comes a confident, full-strength, perfectly crisp wrong answer, indistinguishable forever after from a right one. The insurance pays any claim under the limit and nothing above it. So everything now turns on one question: how often does the hum exceed the limit?

4The barrier comes home

In week 2 a bit was a ball in a valley, a barrier kept it there, and heat occasionally kicked it over — and in Project 1 you tortured that bit yourself: you set the barrier height, you let the heat kick, and you counted the flips. The rule your own chart showed you: raise the barrier and the flips don't just get rarer — they get exponentially rarer. Each notch of barrier multiplied the quiet.

Tonight that model stops being a toy, because a noise margin is a barrier. For the hum to flip a bit in transit, the sloshing on the bucket must climb the full height of the margin in one unlucky moment — exactly the ball's unlucky kick over the wall. And the bookkeeping even matches quantitatively: pushing a bucket's level a whole margin off its rail takes an energy that grows with the margin squared (week 11's square, working for us this time), while the thermal stake funding the attempt is k_BT — the sloshing, squared. Week 2's exponential, e to the minus (barrier over stake), becomes:

Don't memorize the half; keep the shape. The thing in the exponent is the ratio of margin to sloshing — how many times taller the wall stands than the ordinary waves — and it enters squared, which makes this the most violently nonlinear dial in the course. Watch what it does to actual numbers:

And now the tease's question answers itself. How low can the dial go before the sloshing looks like a flip? Lower the supply and — Fig 3.1's last line — the whole ladder shrinks in proportion: the margins fall linearly. But the hum is funded by temperature, not by your supply; it does not shrink at all. So the ratio falls linearly, and the error rate — exponential in the ratio squared — climbs like nothing else you have ever plotted. Notes 11 found an energy sweet spot at 0.68 V and treated low voltage as a speed problem. Tonight it gains a second, harder constraint: a floor. Below some voltage — set not by your cleverness but by how noisy your chip's neighborhoods are — the margins stop being taller than the hum, and no amount of patience makes the answers right. The dial's useful range is pinched from both ends: speed pinches from above (miss the tick), noise pinches from below (miss the truth). Find the floor yourself:

The model: margins are 0.44 of the supply (Notes 06's balanced inverter), the flip chance is eq. (4.1), and the chip makes 10¹⁷ decisions a second. Sloshing of 2 mV is the bare thermal jiggle on a 1 fF bucket (Notes 03); bigger values are whispers and wobble stacked on top. Three things to find. One: at 30 mV of sloshing, creep the dial down from 1.0 V — the verdict leaves SAFE below 0.73 V and hits GIBBERISH by 0.60 V. The whole collapse, from never-event to nonsense, fits inside 130 mV of dial. And notice: Notes 11's energy sweet spot, 0.68 V, is below this floor — on a noisy bus, the cheapest voltage is not a legal voltage. Two: the price of the last 100 mV: from 0.80 to 0.70 V the meter climbs seven orders of magnitude — one flip per quarter-million years becomes one flip every nine days. Three: drop the sloshing to 2 mV, the pure jiggle: now even absurdly low supplies stay SAFE. Thermal noise alone was never the threat on a full-size bucket — it is the stacked whispers and wobble, eating a shrunken margin, that set the floor.

One more connection before we leave the barrier. Notes 11 told you the taps got leakier as barriers shrank: "easy to flip and leaky are the same coin." This section is that coin's other face at the level of whole signals: easy to flip and error-prone are also the same coin. Everything that made the chip fast and cheap — lower barriers, lower swings, tighter margins — moved it closer to the hum. There is no trick that gets speed, energy, and safety; there is only choosing your point on a three-way trade with your eyes open.

5The neighbor's whisper

Now the tease's other question: why does a wire's flip shake its neighbor? Picture two wires running side by side for a millimeter — a bus does this on purpose, sixty-four lanes in tight formation. Notes 09 says each wire is a bucket brigade, around 200 fF of capacitance per millimeter. But capacitance is just "two conductors facing each other," and each wire faces two things: the ground plane below it, and the other wire beside it. So some slice of your wire's bucket — at tight spacing, easily a third — is not between you and ground at all. It is between you and the neighbor. You co-own a capacitor you never asked for.

A capacitor's whole personality is that it resists changes in the gap between its plates. So when the neighbor — call it the aggressor — swings a full volt in a hundred picoseconds, the shared capacitor fights to keep the gap constant, which means it drags your side along. Your quiet wire — the victim — gets yanked off its rail: a glitch, a brief bump that then drains away as your gate's tap pulls the wire back home. How big is the bump? The capacitive handshake splits the aggressor's swing in proportion to ownership: the glitch is roughly the supply swing times the fraction of your bucket the aggressor owns, divided by how firmly your own driver holds the line. A weakly held wire that shares a third of its bucket with an aggressor takes a third-of-a-volt punch — most of the way across a 440 mV margin in one hit, before the jiggle and wobble add their share.

The fixes all fall straight out of that sentence, one each:

• Space them out — coupling falls quickly with distance; every extra wire-width of gap shrinks the aggressor's share of your bucket.
• Hire a bodyguard — route a quiet grounded wire between aggressor and victim (a shield); the whisper now lands on the bodyguard, which doesn't care, and almost none reaches you.
• Hold the line harder — a stronger driver on the victim fights the yank and shrinks the bump; the most vulnerable wire on a chip is one that is barely driven.
• Duck the punch — a glitch only matters if it lands at the moment a receiver is deciding; arrange timing so the bus settles before anyone downstream looks (Notes 10's budgets, now defending against noise too).

The model: supply 1.0 V, margin 440 mV; the neighbor's share of the victim's bucket falls with spacing as 1/(1+spacing); glitch = swing × share ÷ drive. Toy numbers, honest shape. Things to find: at 1× drive, the verdict flips from DANGER to SAFE between 1.2 and 1.3 wire-widths of spacing — real buses genuinely are routed with rules like "one extra gap every N lanes." Then park somewhere dangerous and switch on the shield: the glitch collapses to a few tens of millivolts no matter how close the wires sit. That is why the quiet bodyguard wire is one of the oldest tricks in layout. Last: notice the glitch never fully reaches zero — and remember it rides ON TOP of §4's sloshing. Budgets stack.

6Counting the errors

Notes 12 promised to count the errors, so let us count. The sobering part is never the odds; it is the multiplication. A modern chip makes about 10¹⁷ bit-decisions per second — nearly 10²² per day. Run the numbers on odds that sound absurdly safe: a flip chance of one in a trillion per decision — 10⁻¹², odds no casino could imagine — is 10¹⁷ × 10⁻¹² = a hundred thousand flips every second. Even one in 10²⁰ — odds so long they sound like a physics joke — is a flip every seventeen minutes, almost ninety a day. At computing scale, "astronomically unlikely" and "this afternoon" can be the same number. That is why §4's worked example aimed for ratios near twelve: at 10²² decisions a day, nothing less than a double-digit ratio counts as safe.

And then there is the intruder margins cannot stop. The hum we budgeted for is gentle — millivolt waves, occasionally conspiring upward. But every so often something else arrives: a soft error. A cosmic ray strikes the atmosphere and sprays secondaries; one neutron reaches your chip and dumps a whole bucketful of charge into one node in a single hit. Or an alpha particle pops out of a trace impurity in the chip's own packaging and does the same from inside. This is not a wave lapping at the wall; it is a cannonball. No margin you could afford stops it — and yet nothing is broken: test the chip afterward and it is flawless. The bit simply flipped, once, somewhere, and the machine carried on. Soft errors are rare per bit — but multiply by billions of bits and years of uptime, and every big fleet sees them routinely. It is the reason server memory famously wears a seatbelt.

So the engineer's defenses stack in a ladder, cheapest first:

Rung 3 deserves its one paragraph, because it is a genuinely different philosophy and you will meet it everywhere. The idea, in plain words: store a little extra, so that a flip cannot hide. The simplest version is parity: alongside every stored word, keep one extra bit recording whether the word's count of 1s is even. If any single bit later flips — data or the extra bit, doesn't matter — the count stops matching, and the word confesses on the next read: "I have been tampered with." A few more extra bits, placed cleverly (ECC, error-correcting code), do better than hear the confession — they point at the exact culprit bit, and the memory controller flips it back before your program ever sees it. Server memory does this on every single read, silently, thousands of times a second across a data center; that is the seatbelt. The mathematics of which extra bits to keep is a gorgeous field we leave for another course — the physics point is simpler and it closes tonight's circle: below some error rate, you stop buying more margin and start buying confession instead. The margins keep errors rare; the code makes the rare ones harmless.

7Check yourself

Six questions. Commit to an answer by tapping it — the feedback re-teaches either way.

→Work before Session 23

Session 23 is Monday, November 23. Before then:

• Review these notes. The memory test: name the three rogues and what funds each; say in one sentence why digital beats analog at scale; and say what sets the floor under the master dial — and why the floor moves when the chip gets noisier.
• Ten minutes per bench. On Bench 4.1: set sloshing to 30 mV, predict the floor voltage before you drag, then find it; then measure what the dial's last 100 mV (0.80 → 0.70 V) costs in orders of magnitude. On Bench 5.1: at 1× drive, find the spacing where DANGER flips to SAFE; then park somewhere dangerous and flip the shield on.
• Redo the §4 worked example yourself: take ratio 10 and ratio 5, compute the flip chance from eq. (4.1), multiply by 10¹⁷ decisions per second, and translate each into human time. Bring both numbers Monday.
• Bring one thing: one real-world soft-error story, one sentence. (Famous ones involve an election count, airplanes, and supercomputers — any search for "cosmic ray bit flip" finds the trailheads.)

And the tease. We now know, for one layer of logic, how fast it can go (Notes 10), how cheap it can run (Notes 11), and how safe it can be (tonight). Next week the circuits learn to remember: sequential logic, the clock as referee, and the budget where speed and safety finally collide — timing closure. You will also meet the strangest failure in this course: a flip-flop caught in the act of deciding, hanging between 0 and 1, belonging to neither — metastability, the forbidden zone with a clock pointed at it. And then we climb: from transistors all the way up to the mathematician's idealized machines — finite-state machines, and Turing's. Notes 13 goes from silicon to the top of the stack.

§Symbols and terms