A power grid looks like a network of metal wires, and at low speed it behaves like one. But at the switching rates of modern AI silicon, the grid stops acting like wiring and starts acting like a continuous electromagnetic surface — a medium that stores energy in fields, carries disturbances across the die as waves, and couples every conductor to its neighbors. This is not a modeling convenience. It is what the physics actually does. Here is why, from first principles.
A conductor’s influence is its field — and fields fill space
The starting point is a fact that is easy to forget: a wire does not act on the world through the metal itself. It acts through the fields it sets up in the space around it. A current produces a magnetic field that circles the conductor and extends outward. A voltage produces an electric field that reaches out from the conductor to wherever it can terminate. Neither field stops at the surface of the wire.
That has a direct consequence the moment you place a second conductor nearby: the second conductor is sitting inside the first one’s field. The energy of an electromagnetic interaction is not stored in the electrons inside the metal — it is stored in the fields occupying the space between and around the conductors. So if two conductors are close, their field regions overlap, and the energy in that overlapping region is common to both. That overlap is the entire mechanism behind everything that follows.
Two conductors always share energy — through exactly two channels
When the fields of two conductors overlap, energy can pass between them along two distinct paths:
The magnetic channel — mutual inductance. The magnetic field looping around one conductor’s current threads through the loop formed by its neighbor. If the first current changes, the flux through the neighbor changes, and a changing magnetic flux through a loop induces a voltage in it (Faraday’s law). A current ramp in conductor A therefore drives a voltage in conductor B:
This is transformer action between two windings that never touch — energy shared entirely through the overlapping magnetic field.
The electric channel — mutual capacitance. The electric field reaching out from one conductor’s voltage lands partly on its neighbor instead of only on its own return. A changing voltage then pushes charge onto that neighbor:
This is simply the capacitor that any two conductors with space between them inevitably form.
This is universal, not a power-grid quirk. Any two conductors anywhere have a non-zero mutual inductance and mutual capacitance; it can never be exactly zero. It is the same physics as crosstalk between signal traces, the coupling between transformer windings, or between two antennas. The only variable is strength: fields weaken with distance, so the closer the conductors, the more of one’s field the other intercepts. Closer always means stronger coupling.
A power/ground pair is a transmission line
Now apply this to power delivery. A power rail and its adjacent ground return, carrying equal and opposite current, are not two independent wires — together they are a transmission line: distributed series resistance and inductance along the conductors, and distributed shunt capacitance between them. The pair stores energy in two forms as current and voltage move along it:
and it presents a characteristic impedance set by the ratio of those two stores, Z0 = √(L/C). The deliberate pairing of power with ground does something important to the coupling described above: because the two conductors carry equal and opposite currents, their magnetic fields largely cancel at a distance. A tight power/ground pair therefore radiates far less field into its surroundings than a lone wire would — so it couples less energy into neighboring pairs. This is a deliberate mitigation. It reduces the sharing; it does not eliminate it, because the cancellation is never perfect and the conductors still occupy real space.
A mesh of pairs is a two-dimensional electromagnetic surface
An on-chip power grid is not one transmission line — it is a dense, periodic array of them tiling the die in two dimensions, all electrically connected. Tile a plane with distributed-L–C transmission lines and you no longer have a circuit; you have a medium: a two-dimensional surface that can store energy in its fields and transport disturbances across itself as waves. A voltage disturbance launched at one point travels outward at a finite velocity set by the inductance and capacitance per unit area:
where s is the spacing between pairs and CA the capacitance per unit area. Because the medium has both energy stores, it does what every wave-bearing medium does: disturbances propagate with delay, reflect at boundaries and discontinuities, form standing waves, and resonate where the geometry traps energy bouncing between the inductive and capacitive stores:
Why on-chip silicon actually reaches this regime
None of this would matter if currents changed slowly — the inductive and propagation effects would be negligible and the grid would look like simple wiring. Three facts about modern AI silicon push it firmly into the electromagnetic-surface regime:
| Driver | What it does |
|---|---|
| Extreme di/dt | Current ramps of 10–100 A/ns make the inductive voltage L·di/dt a first-order effect, so the magnetic energy store can no longer be ignored. |
| Heavy on-die capacitance | Large capacitance per unit area makes the wave velocity v = 1/√(LC) slow, which shrinks the effective wavelength of a disturbance. |
| Large electrical size | A slow v plus picosecond-scale rise times means a millimeter-scale die is electrically large: the grid is no longer at one potential at one instant. Propagation delay across the die becomes observable. |
Once the die is electrically large and the inductive store is active, the “everything is instantly at the same voltage” assumption fails. The only description that survives is the field one.
What the surface does that wiring cannot
Treating the grid as an electromagnetic surface is not abstraction for its own sake — it predicts behavior that a wiring picture structurally cannot:
- Distributed, propagating droop. A local current event does not produce an instantaneous global voltage drop. It imposes a voltage surface that spreads across the die over time — a wave, not a step.
- Resonance. Energy ringing between the inductive and capacitive stores produces amplification at specific frequencies. A workload whose spectrum lands on a resonance sees droop far larger than its current alone would suggest.
- Capacitive focusing. Regions of differing capacitance refract and concentrate the propagating disturbance, much as a lens focuses a wave — producing hotspots that are not where the current is.
- Pair-to-pair influence. Because every pair is embedded in the same shared medium, a disturbance under one block reaches and disturbs others — through propagation across the connected surface and, secondarily, through the direct mutual-L/C field coupling between neighbors.
Why we model it as an electromagnetic surface
This is exactly why PI-FP — the engine behind PDNLab — solves the on-chip grid as a two-dimensional distributed transmission-line continuum rather than a network of lumped elements. Only the field/medium formulation captures propagation delay, resonance, capacitive focusing, and the spatially- and temporally-distributed voltage surface that one region’s activity imposes on the rest of the die. The grid behaves as an electromagnetic medium because, at these speeds, energy genuinely lives in and moves through its fields — and modeling it as anything less throws away the very effects that decide whether the supply holds.