Photonic Computing: Using Light Instead of Electrons
Every processor you've ever used moves information as electrical signals through copper and silicon. Photonic computing replaces some or all of that with light — photons traveling through waveguides and optical components etched into silicon. It sounds futuristic, but several companies are shipping real products, and the physics advantages are compelling enough that major tech firms are investing heavily.
The Physics Argument for Photons
Light has properties that electrons don't. Photons travel through waveguides with essentially zero heat generation. Multiple wavelengths can share the same waveguide simultaneously (wavelength division multiplexing) without interfering with each other. And optical signals propagate at roughly 2×10⁸ meters per second in silicon waveguides — about 100× faster than electrical signals in copper interconnects at advanced nodes.
For matrix multiplication — the operation at the heart of neural network inference — photonic circuits can perform the computation in a single pass through an optical network. An electrical processor needs to sequentially multiply and accumulate values; a photonic processor encodes values as light intensities, passes them through an interference network, and reads out the result. The operation completes in the time it takes light to traverse the chip, regardless of matrix size.
There's a fundamental energy advantage too. Moving a bit electrically across a chip at 7nm costs roughly 10-100 femtojoules. Moving a photon the same distance costs an order of magnitude less. When you're doing trillions of operations per second, that difference adds up.
Lightmatter: The VC-Backed Frontrunner
Lightmatter has raised over $400 million to build photonic AI accelerators. Their approach uses Mach-Zehnder interferometers (MZIs) — optical components that split and recombine light beams with programmable phase shifts — to perform matrix operations.
This connects to the ideas in TSMC N2 vs Samsung SF2 vs Intel 18A — So sánh Chip 2nm.
Their Envise chip performs matrix-vector multiplications optically, with digital electronics handling the nonlinear activation functions and data conversion. It's a hybrid approach: photonics for the heavy linear algebra, electronics for everything else.
Lightmatter's Passage interconnect technology is arguably even more interesting than their compute chip. It uses photonic connections between chips in a data center, replacing electrical cables with optical links that offer higher bandwidth density and lower latency at longer distances. For large AI clusters where inter-chip communication is the bottleneck, this could be more valuable than the compute acceleration itself.
Luminous Computing and Analog Photonics
Luminous Computing takes a different approach, aiming for fully photonic transformers that keep data in the optical domain through multiple layers of computation without converting back to electrical signals. This eliminates the optical-electrical-optical (OEO) conversion overhead that plagues most photonic designs.
The technical challenge is enormous. Optical components have limited precision — typically 4-6 bits equivalent, compared to 8-32 bits for digital electronics. Cascading imprecise operations compounds errors quickly. Luminous has published research on optical error correction techniques, but whether they can achieve sufficient accuracy for production LLM inference remains an open question.
For a related perspective, see DDR5 vs LPDDR5X: Memory Architecture, Bandwidth, and Power E.
Silicon Photonics for Interconnects
Even if photonic compute doesn't pan out in the near term, photonic interconnects are already shipping. Intel's been manufacturing silicon photonics transceivers for years. TSMC has a silicon photonics program. Broadcom, Marvell, and Cisco all ship optical networking components.
The near-term application that matters most is co-packaged optics (CPO) — integrating optical transceivers directly into switch and processor packages rather than using pluggable optical modules. CPO reduces electrical trace lengths, cuts power consumption per bit by 30-50%, and enables bandwidth densities that pluggable optics physically can't achieve.
For AI data centers, CPO is probably the first photonic technology that'll become ubiquitous. Moving 100+ TB/s of data between racks through optical fibers is already standard; CPO extends the optical domain right up to the chip boundary.
The Precision Problem
Here's the honest challenge with photonic computing: analog optical components aren't precise. A Mach-Zehnder interferometer's transfer function depends on fabrication tolerances, temperature, wavelength drift, and aging. Maintaining even 8-bit equivalent precision across thousands of optical elements is genuinely difficult.
For a related perspective, see Silicon Wafer Supply Chain: From Sand to 300mm Wafers.
For AI inference, reduced precision (INT8, INT4) is increasingly acceptable, which plays to photonics' strengths. But training — which typically needs higher precision — remains a harder target. Some photonic designs use digital electronics for accumulation and only do the multiply step optically, which helps but limits the efficiency gains.
Temperature sensitivity is another headache. Silicon's refractive index changes with temperature, which shifts the operating point of every optical component. Active thermal tuning consumes power and adds control complexity. Most photonic chips need to operate within a 1-2°C temperature range, which requires either active cooling or sophisticated calibration loops.
Timeline: When Does This Matter?
Photonic interconnects between chips: already here, scaling rapidly. Expect CPO in mainstream data center switches by 2027-2028.
Photonic AI inference accelerators: early production deployments 2026-2027, primarily for specific workloads where the energy efficiency advantage justifies the limited precision and new software stack.
General-purpose photonic computing: likely 2030+, if ever. The control and precision requirements for arbitrary computation in the optical domain are still formidable research challenges.
I'd bet on the interconnect applications being transformational within five years. For compute, photonics will find its niche in energy-constrained inference scenarios, but it won't replace electronic processors for general computing anytime soon. The physics is compelling; the engineering is still catching up.