6.5.15 · D5Advanced & Emerging Architectures
Question bank — Photonic and optical interconnects
True or false — justify
True or false: A photon in a silicon waveguide travels faster than an electrical signal in a good copper trace.
False — light in silicon moves at , often slower per metre than a copper trace's signal velocity, which is already a large fraction of . See Refractive index and speed of light in media.
True or false: Photonics' biggest win is cutting raw latency.
False — the real win is bandwidth density and energy per bit; latency only improves indirectly by removing repeaters and retimers.
True or false: Doubling a copper wire's length doubles its RC delay.
False — it quadruples it, since scales with ; both and grow with , so their product grows with . See Copper interconnects and RC delay.
True or false: A photonic waveguide's transit time also grows like .
False — transit time is , linear in , because light just propagates; nothing is being charged and discharged.
True or false: WDM raises capacity for free because extra colours don't steal each other's bandwidth.
True in the linear sense — since independent colours add — but not free in power or thermal budget; each channel needs its own source/ring. See Wavelength Division Multiplexing (WDM).
True or false: A photodetector outputs a voltage directly, mirroring the modulator.
False — a photodiode outputs a small current ; a TIA is needed to make a usable voltage, and it costs real energy.
True or false: Silicon is a good material for building the laser itself.
False — silicon has an indirect bandgap and emits light poorly; lasers use III–V materials (InP, GaAs) that are bonded or co-packaged. See Silicon Photonics.
True or false: In a Mach–Zehnder modulator, a phase shift of gives maximum brightness.
False — gives zero power (destructive interference, bit 0); gives full power. Recall . See Mach-Zehnder Modulator.
True or false: Energy per bit for a photonic link roughly rises with distance, like copper.
False — once light is launched it is roughly distance-independent, whereas copper's energy per bit rises with distance. See Energy per bit as an efficiency metric.
Spot the error
"Because , we can make bandwidth unlimited by cranking ." — find the error.
is bounded: channels must be spaced apart to avoid crosstalk, and the usable optical band (e.g. C-band) is finite; each channel also draws power and needs thermal tuning.
"Photons don't heat the waveguide at all, so photonic links dissipate zero energy." — find the error.
The transport medium barely heats, but the laser, modulator drivers, and TIA all burn real power — the electrical ends dominate the energy budget.
"A micro-ring resonator blocks all wavelengths equally, acting as a generic shutter." — find the error.
A ring only traps light at one resonant wavelength; that selectivity is exactly why it pairs with WDM (one ring per colour), unlike a broadband shutter.
"Since power is field amplitude, the MZM output is ." — find the error.
Power is the square of the field, ; squaring the sum of two phasors is what produces the interference term.
"Skin effect makes copper resistance fall at high frequency, helping speed." — find the error.
The opposite — current crowds to the surface, so effective resistance rises roughly with , worsening loss at high data rates.
"E-O-E means the signal stays optical end to end." — find the error.
E-O-E is Electrical→Optical→Electrical: the data starts and ends as electrons; light only carries it through the middle.
Why questions
Why does WDM add capacity linearly instead of sharing one channel in time like an electrical bus?
Different colours occupy different wavelengths and don't interfere, so they transmit simultaneously; a shared electrical bus must time-slice one medium, dividing rather than adding capacity.
Why is a carrier a plus when we only modulate at a few GHz?
A few-GHz data wiggle is a tiny fraction of a optical carrier, leaving enormous frequency headroom — the carrier is nowhere near saturated.
Why does interference let a Mach–Zehnder do "logic"?
Splitting light and recombining after an electrically-set phase shift makes the two arms add (bright = 1) or cancel (dark = 0), so a voltage controlling literally switches the bit. See Mach-Zehnder Modulator.
Why does co-packaged optics help escape the "pin wall"?
Putting the laser and modulator right next to the CPU/GPU package lets bandwidth leave on fibre/waveguides instead of a limited number of electrical package pins. See Co-packaged optics.
Why is an on-chip photonic NoC still "emerging" rather than mainstream?
Integrating lasers into CMOS and keeping ring resonators on their target wavelength via thermal tuning are hard, power-hungry problems that aren't yet cheaply solved.
Edge cases
What is the MZM output at ?
— the halfway "quadrature" point, useful as a bias but ambiguous as a clean 0 or 1.
What happens to a micro-ring's behaviour if the chip heats up and its resonance drifts?
Its resonant wavelength shifts off the intended channel, so it stops selecting the right colour — which is why active thermal tuning is required to hold resonance.
In the limit of a very short copper wire, does photonics still win?
Not necessarily — at short lengths delay is tiny and the fixed cost of laser + modulator + TIA dominates, so copper can be cheaper and lower-energy for short links.
What is the photodetector current when zero light arrives (a "0" bit)?
Ideally , but real diodes leak a small dark current, so the receiver must threshold above that floor to read a clean 0.
At , how does the MZM output compare to ?
Identical — , full brightness; the response is periodic in , so the drive voltage only needs to reach a swing to toggle.
Recall One-line summary to carry away
Photons win on bandwidth density and energy, not on raw metres-per-second; every claim about "unlimited" or "faster" hides a boundary — crosstalk, finite band, indirect bandgap, or the electrical ends that still burn the power.