The Interconnect Bottleneck

The last module ended with a defeat: optical transistors lost to CMOS, comprehensively. This module opens with the counter-attack — and it starts from a number every computer architect should be able to recite. In a modern process, a 32-bit addition costs roughly 0.1 pJ. Fetching those same 32 bits from off-chip DRAM costs roughly 640 pJ — six thousand times more. Even hauling them a few millimetres across the die, at around 100 fJ per bit per millimetre, costs more than computing with them. Arithmetic is nearly free; data movement is the bill. You met this asymmetry as the memory wall, and its physical root as the interconnect wall: transistors kept shrinking beautifully, and wires did not.

Photonics does not need to out-switch the transistor to matter. It needs to out-carry the copper wire — and at that job, light is not merely competitive; beyond a certain distance it is the only game in town. This lesson is about locating that distance, and understanding the physics that pins it where it is.

Why copper runs out of breath

A short on-chip wire is an RC diffusion line: the signal doesn't propagate down it like a wave so much as ooze, charging the wire's distributed capacitance through its distributed resistance. Double the length and you double both R and C — delay grows as the square of length, which is why long on-chip routes are chopped up with repeaters, each burning energy of its own. Worse, scaling makes it relatively harder every generation: a thinner wire has more resistance per millimetre, so while transistors sped up, a fixed millimetre of wire stood still.

Off the chip, on a board or a cable, the wire becomes a proper transmission line — but a lossy one, and the loss grows with frequency. Two mechanisms gang up:

Skin effect — at high frequency, current crowds into a thin skin of depth \delta = \sqrt{2\rho/\omega\mu} at the conductor surface, so the effective resistance (and attenuation, in dB) grows like \sqrt{f};
Dielectric loss — the board material absorbs energy each cycle, adding attenuation that grows like f, and dominating at the highest rates;
together they make a copper channel a steep low-pass filter: a fast edge launched at one end arrives smeared, its high-frequency content preferentially eaten;
the practical consequence is a roughly constant bandwidth × distance product — a channel that carries 200 Gb/s over one metre manages only ~100 Gb/s over two.

Engineers fight the filter with equalization: pre-distort the transmitted waveform, then run the received one through adaptive filters and DSP that undo the smearing. It works — a modern 224 Gb/s SerDes recovers an eye from a channel with over 30 dB of loss — but every filter tap and every DSP operation costs energy, charged per bit, forever. The pJ/bit of an electrical link is mostly the price of arguing with the channel. And here is the contrast that powers this whole module: an optical fibre attenuates about 0.2–0.4 dB per kilometre, and that figure is essentially independent of the data rate, because a 100 GHz signal is a trifling fractional bandwidth on a 200 THz carrier. Copper's loss grows with both distance and speed; glass's grows with neither, at any rate you can afford to modulate.

The crossover

Put the two technologies on one chart. An electrical link's energy per bit starts small but climbs with distance — longer wires mean more capacitance to charge, more loss, more equalization. An optical link pays a large fixed toll up front — laser, modulator, photodetector, receiver — and then almost nothing per additional centimetre. Two straight-ish lines with different slopes must cross:

Everything strategic about optical interconnects lives in that crossover point. In the 1980s it sat at kilometres, and optics took long-haul telecoms. By the 2000s it reached a few hundred metres, and optics took the datacenter's longer links. Today it hovers around a metre or two and is still marching inward: each time the fixed toll of conversion drops — better modulators, integrated lasers, co-packaging — light wins another ring of territory around the chip. Slide the toll down and watch the conquest. The rest of this module is the story of that march, ring by ring, from pluggable transceivers to co-packaged optics to the contested interior of the chip itself.

The beachfront problem

Energy is one wall; geometry is the other. Every bit entering or leaving a chip must cross the chip's edge (or, with modern packaging, its perimeter and surface). Architects call this the shoreline or beachfront problem: a die is an island, and you can only build so many harbours per millimetre of coast. A big switch ASIC is about 25 mm on a side — 100 mm of shoreline. State-of-the-art electrical SerDes escape roughly 500 Gb/s per millimetre of die edge. Multiply:

100\ \text{mm} \times 0.5\ \text{Tb/s/mm} = 50\ \text{Tb/s}

— which is almost exactly the aggregate bandwidth of today's 51.2 Tb/s switch chips. That is not a coincidence; it is a design sitting at the wall. Meanwhile switch bandwidth has doubled every two years for a decade, and the shoreline hasn't grown a millimetre. Run the collision forward yourself:

const edgeMm = 25; // die edge, mm const shoreline = 4 * edgeMm; // perimeter, mm const electricalGbsPerMm = 500; // escape density, electrical SerDes const opticalGbsPerMm = 2000; // escape density, dense optical I/O (fibre + WDM) const escapeElectrical = shoreline * electricalGbsPerMm / 1000; // Tb/s const escapeOptical = shoreline * opticalGbsPerMm / 1000; console.log("electrical escape: " + escapeElectrical + " Tb/s"); console.log("optical escape: " + escapeOptical + " Tb/s"); // Switch bandwidth doubles roughly every 2 years. Starting at 51.2 Tb/s: let demand = 51.2; for (let year = 2024; year <= 2032; year += 2) { const fits = demand <= escapeElectrical ? "fits electrically" : "needs optics (or better)"; console.log(year + ": " + demand + " Tb/s -> " + fits); demand *= 2; }

Optics raises the escape density for two reasons you already own: a single-mode fibre core is a few microns across yet carries terabits via wavelength parallelism — many independent channels sharing one physical port — and its reach is unlimited at datacenter scale, so the harbour doesn't also have to fund a navy of equalizers. More bits per millimetre of beach, at fewer picojoules each: that is the entire pitch.

The folk explanation — "optical interconnects are faster because light travels faster than electricity" — is wrong on both ends. Signals on a copper transmission line already propagate at roughly half the vacuum speed of light, and light in glass travels at about two-thirds c — the two are within a factor of ~1.3, and for latency the fibre is barely better (sometimes worse, once serialization and DSP delays are counted). Nobody deploys optics for propagation speed. The wins are elsewhere entirely: attenuation that is tiny and rate-independent, bandwidth density through a few-micron core carrying many wavelengths, immunity to crosstalk and EMI, and consequently far better energy per bit at distance. When you hear "light is fast", reach for this lesson's chart instead: the correct slogan is "light is cheap to carry far".

If the channel eats your signal, why not simply transmit more power? Because the loss is exponential in distance: a channel with 30 dB of loss delivers one thousandth of the launched power, 60 dB one millionth. Shouting is a losing arms race — every 3 dB of extra loss doubles the required transmit power, and the receiver's own thermal noise floor sets a hard minimum it must clear. The actual escalation path of electrical signalling has been cleverness, not volume: multi-level signalling (PAM4 — two bits per symbol, at the price of a smaller eye), forward error correction (accept errors, then spend logic correcting them), and ever-heavier DSP equalization. Each trick buys a generation and adds picojoules. The 224G SerDes at the end of this road is a genuine marvel of engineering — a little software-defined radio in all but name — and its very sophistication is the measure of how hostile the copper channel has become. Glass simply declines to fight: at 0.2 dB/km, "louder" was never needed.

Where this goes next

So the battleground is drawn: beyond the crossover distance, photons carry bits for less energy and through less beachfront than electrons — if you can pay the conversion toll at each end. That toll is a real machine: a laser, a modulator, a wavelength multiplexer, kilometres of glass, a photodetector and the electronics to read it. The next lesson opens that machine up — the anatomy of a WDM link, and the decibel bookkeeping that decides whether it closes.