Put your hand on a laptop that's working hard. That warmth is the sum of billions of tiny, identical events: a logic gate pulling one wire from 0 volts up to the supply voltage, or letting it fall back down. We have spent this course near the thermodynamic floor — Landauer's principle and its kT \ln 2 per erased bit. This module climbs back up to real silicon and asks the engineer's question: where does the energy actually go when a CMOS gate switches — and how far above the floor are we living?

The answer is startling on both counts. The loss per switch follows one tidy formula, \tfrac{1}{2}CV^2 — and it sits four to five orders of magnitude above Landauer's limit. Almost all the energy your computer burns has nothing to do with the physics of information. It is an implementation habit. This lesson derives the habit; the rest of the module breaks it.

Charging a wire from a fixed rail

In a CMOS gate, driving an output high means connecting the output node — a little capacitance C made of wire and transistor gates — through a pMOS transistor to the fixed supply rail at voltage V. The transistor's channel is not a perfect wire; it has some resistance R. Current rushes through R into C until the node reaches V. Now do the accounting:

• Charge delivered: the capacitor ends up holding Q = CV.
• Energy drawn from the supply: every coulomb left the rail at the full voltage V, so the supply paid E_{\text{drawn}} = QV = CV^2.
• Energy stored on the capacitor: the capacitor filled up gradually — the first charge arrived when the node was near 0, the last when it was near V — so it stores only the average, E_{\text{stored}} = \tfrac{1}{2}CV^2.
• The difference, CV^2 - \tfrac{1}{2}CV^2 = \tfrac{1}{2}CV^2, was burned as heat in the resistance.

The independence from R is the point to sit with. You cannot engineer this loss away with better transistors: make the channel resistance ten times smaller and the current spikes ten times higher, and the I^2 R heat comes out exactly the same. The waste is baked into the method — slamming a capacitor onto a fixed-voltage rail — not into the components. That observation is the seed of the whole adiabatic-circuits idea in the next lesson.

The parabola of pain

The formula is a parabola in V: halve the supply voltage and the switching energy falls by four. That quadratic lever powered decades of progress — supply voltages slid from 5 V in the 1980s to under 1 V today. Drag the capacitance slider to see the scale of a single modern gate.

Real numbers: a gate's node capacitance is around C \approx 0.1–1\ \mathrm{fF} and supplies sit near V \approx 0.7\ \mathrm{V}. That gives

so each switch costs roughly 10^{-16}–10^{-15} joules — a tenth of a femtojoule to a femtojoule. Compare the Landauer floor at room temperature, kT \ln 2 \approx 3 \times 10^{-21}\ \mathrm{J}: we are paying about 10^{4}–10^{5} times the fundamental minimum, per switch — and that is before counting the leakage current that trickles through every transistor even while it does nothing at all. There is a factor of ten thousand lying on the table.

Why not just keep lowering V?

The quadratic lever tempts an obvious plan: keep shrinking V until the parabola meets the floor. It doesn't work, for two compounding reasons:

• The threshold. A MOS transistor only turns on convincingly once its gate passes a threshold voltage of a few hundred millivolts; below it the transistor doesn't switch off — it merely leaks exponentially less. Squeeze the supply toward the threshold and gates become slow and leaky at the same time.
• Noise margins. A logic level must be distinguishable from thermal noise. The natural noise scale is the thermal voltage kT/q \approx 26\ \mathrm{mV} at room temperature, and a reliable digital swing must stay many times above it — tens of kT/q — or bits start flipping themselves, for the same Boltzmann-factor reasons that keep a stored bit in its well. In practice supplies have stalled around 0.7–1 V.

So the conventional road ends: V is pinned from below by physics and reliability, C is pinned by wire lengths and transistor sizes, and \tfrac{1}{2}CV^2 per switch is what you pay. Escaping the parabola will require abandoning its premise — the fixed rail itself.