TSMC bets on unorthodox optical tech

Electrons vs photons and fundamental limits

• Several comments contrast electrons (fermions) with photons (bosons): electrons strongly interact and obey Pauli exclusion, photons mostly pass through each other and interact weakly.
• This makes electrons well suited for logic and nonlinear devices (transistors), while photons are better for high‑bandwidth transport.
• Optical links still have limits: attenuation, noise (OSNR/SNR), and nonlinear effects in fiber at very high powers/bit‑rates, but photon–photon interactions are negligible at the scales discussed here.

Signal integrity: copper vs fiber

• Copper links are limited by signal integrity: interference, attenuation, impedance mismatches, and inter‑symbol interference.
• Fiber has far lower attenuation over distance and supports dense wavelength multiplexing, but suffers from chromatic and modal dispersion; for imaging fibers and multimode links, mode dispersion is a key concern.
• Vibration‑induced phase noise is argued to be irrelevant for intensity‑modulated LED links at these scales.

MicroLED approach vs laser/VCSEL optics

• The discussed tech uses microLED arrays into relatively large‑core fiber bundles (∼50 µm) and CMOS detector arrays.
• Claimed advantages over conventional laser/VCSEL links: significantly lower energy per bit, simpler electronics (no heavy DSP/SerDes), easier coupling/packaging, and potentially better reliability and cost for short reaches.
• Skeptics question whether microLEDs truly beat VCSEL arrays in cost, coupling, and reliability, and note that similar parallel VCSEL+multicore fiber approaches already exist.

Scope, distances, and use cases

• Intended distance is sub‑10 m: intra‑rack or near‑rack links, possibly chip‑to‑chip or board‑to‑board interconnects (PCIe/NVLink/HBM‑class buses), not long‑haul or typical intra‑datacenter runs.
• For longer distances (10 m–km), commenters agree lasers remain necessary.

SerDes, parallelism, and protocol

• Even with 10 Gb/s per fiber, electronic logic runs slower and must serialize/deserialize, but SerDes can be placed at different points along the electro‑optical chain.
• Parallel optics does not remove skew issues entirely but can manage them with equal‑length bundles and per‑lane clock recovery; some propose dedicating “pixels” to timing/control.

Optical computing and neuromorphic ideas

• Commenters reiterate that all‑optical transistors and general photonic CPUs are blocked by weak optical nonlinearities; high intensities needed are impractical.
• Optical neuromorphic and matrix‑multiply accelerators are active areas, but nonlinear activations and training (backprop) remain major obstacles.

Quantum computing optics vs this work

• Quantum platforms need coherent, narrow‑linewidth lasers and often single‑photon or entangled states; incoherent LEDs cannot substitute.
• Some see LED‑based interconnects as orthogonal to, not indicative of failure of, laser‑integrated optics for quantum systems.

TSMC’s role and article framing

• Multiple comments say the headline overstates TSMC’s “bet”; they view it more as a foundry engagement plus some custom detector development.
• Others argue that TSMC doing custom photodetectors at all is itself a meaningful vote of confidence in the technology.