FMCW Array Limitation Analysis

Silicon-photonics vendors routinely describe a “laser on a silicon chip.” The phrase rewards careful reading, because it sits at the center of how — and how well — a coherent FMCW array can scale.

The Blunt Answer: No, Silicon Cannot Lase at 1550 nm

Silicon is an indirect bandgap semiconductor. This is not a manufacturing limitation or an engineering challenge to be engineered around — it is a fundamental consequence of silicon's crystal band structure.

Why Indirect Bandgap Matters

In a direct bandgap material (like InP or GaAs), the conduction band minimum and valence band maximum occur at the same crystal momentum (k-vector). An electron can recombine with a hole and emit a photon directly — the transition conserves both energy and momentum simultaneously, and light emission is efficient.

In silicon, the conduction band minimum and valence band maximum are at different k-vectors. For an electron to recombine radiatively, it must simultaneously change both its energy and its crystal momentum. A photon cannot carry significant momentum — so this transition requires the involvement of a phonon (a lattice vibration quantum) to conserve momentum. This three-body process (electron + hole + phonon → photon) is extremely improbable compared to the competing non-radiative recombination paths. The result is that silicon's radiative efficiency is roughly four to six orders of magnitude lower than direct bandgap III-V materials. You cannot make a silicon laser in any practical sense using bulk crystalline silicon.

This is not a new or contested finding. It has been known since the 1950s. The Nobel Prize-winning work on semiconductor lasers was done with III-V materials precisely because silicon was understood to be unsuitable.

So How Do Vendors Put a "Laser on a Silicon Chip"?

This is where the marketing language becomes technically slippery and deserves careful reading.

Vendors describe creating "lasers on chips" through a fabrication partner. The key phrase buried in such technical descriptions is hybrid integration. What this almost certainly means in practice is one of these approaches:

Option 1: III-V Die Bonding onto Silicon

A separately grown III-V laser chip (indium phosphide, InGaAsP, or similar — materials that do lase efficiently at 1550 nm) is bonded onto the silicon photonics wafer. The laser light is then evanescently coupled from the III-V gain region into silicon waveguides on the chip. The silicon handles routing, modulation, splitting, and detection — but the gain comes from the III-V material, not the silicon.

This is well-established technology. Mature silicon-photonics products for data centres use exactly this approach — III-V lasers bonded to silicon. It works, but it introduces:

• A bonding interface that is a potential reliability weak point
• Thermal mismatch between III-V and silicon that stresses the bond over temperature cycling
• Additional fabrication complexity and cost
• A component (the III-V die) whose supply chain and yield are independent of the silicon fab

Option 2: Flip-Chip Attachment

The laser is a separate packaged component, attached to the silicon photonics chip via flip-chip bonding and coupled through a grating coupler or edge coupler. This is even more of a hybrid — the laser is really a separate device that happens to be attached to the same substrate. Coupling efficiency and alignment stability over temperature are the key engineering challenges.

Option 3: Germanium-on-Silicon

Germanium can be grown on silicon and has better optical properties than silicon. Ge-on-Si photodetectors are well established. There has been research on GeSn (germanium-tin alloy) lasers that could in principle be integrated on silicon, but as of the current state of the art, GeSn lasers are laboratory demonstrations operating at cryogenic temperatures — not production-ready room-temperature devices. This is almost certainly not what these LiDAR vendors are using.

Why This Matters for the Claims Being Made

When a vendor says they have integrated "all photonics functions including lasers" on a single silicon chip, this statement needs to be read carefully. It is almost certainly true that all the passive and active waveguide functions — splitting, routing, phase shifting, modulation, mixing, and detection — are on silicon. These are things silicon photonics genuinely does well. The germanium-on-silicon photodetector is standard and mature. Silicon modulators based on the plasma dispersion effect are standard.

But the laser source is almost certainly a III-V gain element bonded or attached to the silicon — not a silicon laser. The distinction matters for several reasons:

Reliability — the heterogeneous bonding interface between III-V and silicon is a potential long-term reliability concern under automotive thermal cycling. III-V and silicon have different coefficients of thermal expansion. Repeated cycling over -40°C to +85°C stresses the bond. This is a known concern in automotive photonics and one reason why achieving AEC-Q qualification for silicon photonics modules with bonded III-V lasers is non-trivial.

Chirp control — the linewidth and chirp characteristics of the laser are determined by the III-V gain medium, not the silicon platform. Silicon photonics provides waveguides and modulators to help tune the chirp, but the fundamental laser physics — including the linewidth enhancement factor, the response to current injection, the temperature dependence of lasing wavelength — are properties of the III-V material. Calling it a "silicon laser" obscures who actually owns that performance.

Manufacturing — if the laser is a III-V component bonded to silicon, the chip is not manufactured in a standard silicon CMOS fab. It requires a hybrid process involving a separate III-V fab and a bonding step. This is more complex and more expensive than a pure silicon process, and the economies of scale argument — "we use the same fabrication techniques as computer chips" — applies only to the silicon portion of the chip.

The Honest Summary

Silicon fundamentally cannot lase at 1550 nm due to its indirect bandgap. This is physics, not engineering. The laser source in these systems is almost certainly a III-V gain element (likely InP-based) that is heterogeneously integrated onto the silicon photonics platform. The silicon handles the coherent optics — waveguides, couplers, detectors, mixing circuits — and does this genuinely well. But the "laser on a chip" framing in the marketing elides a critical technical distinction.

This doesn't make what they're doing unimpressive — hybrid III-V on silicon is technically demanding and commercially important. But it does mean:

• The system is not a pure silicon device
• The long-term reliability of the III-V/silicon interface under automotive conditions is a real open question
• The chirp characteristics are governed by III-V laser physics, which brings back all the nonlinearity and temperature sensitivity concerns discussed earlier
• The manufacturing cost argument based on CMOS scale economics applies only partially

When you hear "silicon photonics LiDAR chip with integrated laser," the technically precise statement is: a silicon photonics platform for coherent LiDAR with a heterogeneously integrated III-V laser source. The difference between those two descriptions is exactly where several of the hardest unsolved problems live.

A related question decides whether any of this scales: how many laser sources does a wide field of view actually require, and how do they combine? It is one of the most underappreciated problems in FMCW LiDAR.

First: What Determines How Many Laser Sources You Need?

In FMCW LiDAR, each laser source illuminates one spatial channel at a time. Unlike a flash LiDAR where you flood the scene, FMCW fundamentally works by mixing a local oscillator copy of the outgoing chirp with the return from a single spatial direction. The coherent mixing only works when the local oscillator and the return are from the same laser, same chirp instance.

So the question of how many lasers you need is really the question of: how do you cover a large field of view with enough points, fast enough?

The Field of View Problem — The Numbers

Consider automotive-grade requirements:

• Horizontal FoV: 120° front, ideally 360° surround
• Vertical FoV: ~30° (roughly -15° to +15°)
• Angular resolution: 0.1° × 0.1°
• Update rate: 20 Hz full scene

At 0.1° resolution over 120° × 30°:

Total points per frame = (120/0.1) × (30/0.1) = 1,200 × 300 = 360,000 points

At 20 Hz: 7.2 million points per second

Now, how fast can a single FMCW channel acquire one point? A single chirp sweep needs to be long enough to give you adequate range resolution. Range resolution in FMCW is:

ΔR = c / (2B)

where B is the chirp bandwidth. For 15 cm range resolution you need B ≈ 1 GHz. A chirp sweep of 1 GHz at a reasonable sweep rate takes on the order of 1–10 microseconds per point, depending on your SNR budget and processing approach.

Call it 2 μs per point conservatively. Then a single laser channel can acquire:

1 / 2×10⁻⁶ = 500,000 points per second

But you need 7.2 million points per second for the full front FoV at 20 Hz. That means:

7,200,000 / 500,000 ≈ 14–16 parallel channels minimum

For 360° surround at the same resolution and update rate, multiply by 3: ~48 parallel channels.

And this is the optimistic case — it assumes perfect chirp duty cycle with no dead time between sweeps, no time lost to beam settling, and no margin for processing overhead.

So You Need Multiple Laser Sources — How Do They Stack?

There are three fundamentally different architectures for handling multi-channel FMCW, and they have very different implications for complexity, cost, and the problems we've been discussing.

Architecture 1: One Laser, Time-Multiplexed Across Channels

A single laser is split and its output steered sequentially to different spatial positions — one at a time.

Single laser → beam splitter → steering mechanism, addressed sequentially: point 1 → point 2 → point 3 → …

Advantage: One laser, one chirp source, one set of chirp linearity problems to solve.

Disadvantages:

• You are limited to sequential acquisition — the scan rate bottleneck is severe
• To cover 360,000 points at 20 Hz with one laser channel, each point gets only 138 nanoseconds of dwell time. At 1550 nm with realistic return signal levels, that is not enough time to accumulate the photons needed for adequate SNR at 200 m range
• Essentially forces you into a mechanical scanning architecture to get any reasonable point density
• One laser failure = total system failure

Verdict: Fine for short-range industrial applications with modest point density requirements. Not viable for automotive-grade surround LiDAR.

Architecture 2: Multiple Independent Lasers, Each with Its Own Chirp Source

Each spatial channel has its own laser and its own chirp generation electronics. They operate in parallel, each covering a portion of the FoV.

• Laser 1 + chirp gen → channel 1 (angular sector 1)
• Laser 2 + chirp gen → channel 2 (angular sector 2)
• Laser 3 + chirp gen → channel 3 (angular sector 3)
• … up to Laser N + chirp gen → channel N (angular sector N)

Advantage: True parallel acquisition. Each channel has full dwell time. Straightforward to scale.

Disadvantages — and this is where it gets serious:

Each laser source needs its own chirp generation and linearisation. If you need 16 channels, you need 16 independent chirp sources, each with sub-MHz linewidth, each generating a linear GHz-bandwidth chirp, each characterised and compensated for temperature drift — independently.

The chirp linearity problem we discussed does not get easier when multiplied by 16. It gets 16 times harder to manage. And critically:

The chirps from different lasers must not interfere with each other. If laser channel 3's return signal gets mixed with the local oscillator of channel 7, you get a false range reading. Managing inter-channel crosstalk in a coherent system where all lasers are operating at similar wavelengths near 1550 nm requires either:

• Frequency-division multiplexing — each laser operates at a slightly different centre wavelength, and narrowband filters separate them at the detector. But the wavelength spacing must be larger than the chirp bandwidth, which pushes you into managing a set of wavelength-specific components
• Time-division multiplexing between nearby channels — which partially defeats the parallelism advantage
• Physical isolation — spatial separation of channels so their beams don't overlap on the detector. This works in principle but requires careful optical design

Verdict: Scalable in principle, but multiply every chirp linearity and temperature stability problem by the channel count. Cost and complexity scale roughly linearly with channel count. A 16-laser system with 16 independent chirp sources is a formidable engineering challenge.

Architecture 3: One Master Laser, Split into Multiple Parallel Coherent Channels

One high-quality laser source is used, and its output is split into multiple copies — one for each channel. Each copy goes through its own modulator to impose the frequency chirp, and each modulated copy serves one spatial channel.

• Master laser → splitter → modulator 1 → channel 1 detector
• Master laser → splitter → modulator 2 → channel 2 detector
• Master laser → splitter → modulator 3 → channel 3 detector
• … and so on, one branch per channel

Advantage: The laser coherence length and linewidth are shared across all channels — you only solve that problem once. Inter-channel coherence is guaranteed because all channels derive from the same source. This also means the local oscillator for each channel is inherently phase-coherent with the transmitted signal, which is exactly what coherent detection requires.

Disadvantages:

• Each split reduces the optical power per channel by 1/N. With 16 channels, each channel gets 1/16 of the laser power. At the return signal levels involved in 200 m LiDAR, this is a serious SNR problem. You need either a very powerful master laser — which raises eye safety and thermal management issues — or optical amplifiers (SOAs — semiconductor optical amplifiers) in each channel to restore power
• SOAs add noise (amplified spontaneous emission, ASE), add cost, add power consumption, and add more components that can fail
• The modulator in each channel must generate an accurate, linear frequency chirp — the chirp nonlinearity problem reappears at the modulator level. You've traded one laser chirp problem for N modulator chirp problems
• Silicon photonics modulators based on plasma dispersion are inherently nonlinear and have bandwidth limitations that make broadband chirp generation non-trivial

Verdict: The most architecturally coherent approach for a chip-integrated system of this kind. But it introduces a power splitting tax that gets severe at high channel counts, and moves the chirp nonlinearity problem from the laser to the modulators — where it is somewhat more tractable (modulators are more amenable to electronic pre-distortion) but not eliminated.

How Channels Stack on a Silicon Photonics Chip

This is where silicon photonics has a genuine advantage — and also where the real constraints become visible.

On a silicon photonics chip, channels can be implemented as parallel waveguide paths. Each channel has its own:

• Splitter tap from the master waveguide
• Phase modulator or electro-optic modulator for chirp generation
• Output grating coupler or edge coupler (the interface to free space)
• Return input coupler
• Balanced photodetector pair (for coherent detection)
• 90-degree optical hybrid (for I/Q detection)

A 4-channel chip is feasible and demonstrated — commercial short-range modules appear to implement something in this range. Scaling to 16 channels on a single chip is architecturally possible but faces:

Die size — each channel adds area. Silicon photonics dies are not free to scale arbitrarily. A 16-channel coherent LiDAR chip would be a large die with corresponding yield implications.

Power dissipation — 16 channels of modulators, amplifiers, and detectors generate heat that must be managed. Thermal gradients across the chip affect the refractive index of silicon waveguides (silicon has a strong thermo-optic coefficient, dn/dT ≈ 1.8 × 10⁻⁴ /°C), which shifts the phase in every waveguide on the chip — including the local oscillator paths. This creates a thermally induced chirp error that couples back into the range measurement. On a large multi-channel chip, differential thermal gradients between channels are particularly problematic.

The OPA coupling problem — if each channel is also connected to an OPA for beam steering, the number of phase-controlled antenna elements multiplies dramatically. A 16-channel system where each channel drives a 64-element OPA requires 1,024 independently controlled phase elements, each needing sub-wavelength precision across temperature. The control electronics for this alone are a significant system engineering challenge.

The Honest Summary of the Scaling Problem

The core tension is this: everything that makes FMCW work well — coherent detection, high SNR from long integration, immunity to interference — assumes a clean, linear, stable chirp. Every additional channel you add to cover a larger field of view is another instance of the chirp control problem, another source of thermal cross-coupling, and another component in the reliability chain.

A 4-channel chip for a short-range industrial scanner is an impressive engineering achievement. Scaling the same architecture to 48 channels for automotive surround LiDAR is not a linear extrapolation — it is a qualitatively different engineering problem. And as far as the public technical record shows, nobody has solved it yet.

This is precisely why the new paradigm argument matters so much. An architecture that eliminates the per-channel chirp generation problem — through scene-wide illumination, through a fundamentally different ranging modality, or through a different approach to parallelism — sidesteps a scaling wall that FMCW chip-integrated approaches will inevitably hit.