Photonics' Third Wave: Out of the Lab in Five Years, Conquering Drones First
Photons cannot interact like electrons, but wavelength multiplexing enables massive parallelism and ultra-low latency. All-optical neural networks could land within five years, with the first stop not servers but drones and real-time control.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Photons don't interact, yet they compute in parallel
Photons do not interact with each other as easily as electrons, which has long been the reason photonics is used for communication rather than computation. But Patti points out that this 'weakness' brings two major advantages: extremely high parallelism and very high speed. Wavelength Division Multiplexing (WDM) packs multiple light signals of different colors into the same fiber, directly multiplying the information throughput a single chip can handle. In other words, photons can compute, just in a different way—using parallelism to compensate for the lack of interaction.
— Patti StabileAll-optical neurons have been realized
Patti's team has achieved an 'all-optical' implementation: neurons, inter-layer connections, and nonlinear activations are all performed in the optical domain using semiconductor optical amplifiers. Signals travel from the first layer's input to the last layer's output without ever returning to the electrical domain or accessing intermediate memory. This breaks the limitation of earlier designs where photonics only handled inter-layer connections while neurons still required electronic implementation. The biggest benefit of all-optical computing is ultra-low latency—light traversing a chip takes far less time than electronic circuits accessing memory.
— Patti StabileNoise compression in optical amplifiers is a hidden advantage
The nonlinear activation of semiconductor optical amplifiers not only performs the threshold function but also has a often-overlooked role: compressing noise and regenerating signals. This means that as network size grows, accuracy does not degrade layer by layer—a key to scalability, not just 'fast computation.' The downside is size and power consumption. Patti admits that amplifiers are current-driven and power-hungry; switching to voltage-driven elements like microring resonators could reduce power consumption by one to two orders of magnitude, which is already in the plans.
— Patti StabileProducts within five years, the third wave
Patti believes photonic computing is in its third wave. The first wave has long passed, the second receded, and recent AI compute demand has rekindled interest. Her prediction: real product prototypes will appear within three to five years. The first stop will not be replacing large-scale compute engines in data centers, but real-time control scenarios like drones and autonomous driving that require ultra-low latency, lightweight, and low power. Optical signals enter the chip from sensors, inference is performed directly in the optical domain, and the next action is driven—the entire loop completes in an extremely short time.
— Patti StabilePhotonics is the communication backbone for neural networks
Ralph offers a key judgment: photonics' greatest value lies not in fully replacing electronic computation, but in the communication backbone of neuromorphic networks. He recalls David A.B. Miller's famous quote that 'the brain can fit in a pizza box,' and argues that optical interconnects are the core driver for scaling up chip-to-chip communication. He questions the necessity of all-optical computing: rather than pursuing optical implementation of all layers, it might be more transformative to let photonics handle data transmission between boards and racks, embedding only a small amount of neural processing.
— Sunny BainesThe power budget only works when scaled up
Ralph clarifies a commonly misunderstood power logic: photonic systems are not more efficient than electronics at the component level, but system-level efficiency improves with scale. The initial laser injection is a fixed cost; the larger the computation, the lower the per-operation power after amortization, and accuracy is also higher. Doing photonics at small scale is a disaster; only at large scale can it become an advantage. However, he cautions that photonics lags the overall neuromorphic engineering field by about a decade, and the infrastructure and toolchain are not yet mature.
— RalphDon't compare photonics to traditional digital
Ralph sharply criticizes the comparison approach in the literature: if photonic chips are only compared to 'off-the-shelf traditional digital architectures' in terms of power and speed, that is an unfair win. The real comparison should be against analog and digital neuromorphic approaches, which are also emerging technologies. Giulia adds a more concrete example: MIT's photonic processor for 6G wireless signal classification is 100 times faster than digital solutions, with a latency of only 120 nanoseconds—but Ralph points out that this actually proves photonics excels at communication, not computation per se.
— RalphIn their own words · checked verbatim
I don't think the problem are the laser per se because again, also they don't scale linearly as long as they encode information into your chip.
Patti Stabile15:25
It's not necessarily that it's de facto each component is more efficient, but if you're using the system right, the system as a whole can be more power efficient.
Ralph36:32
most of optical computing was really not optical computing, most of it was optical communications within computing.
Sunny Baines39:32
You should not be comparing yourself with that. You should be comparing yourself with the other emerging technologies.
Ralph42:33
MIT did a photonic processor that could streamline 6G wireless signal processing... running 100 times faster than the digital alternatives in 120 nanoseconds.
Giulia D'Angelo43:33
Figures
| Patti's predicted timeline for photonic computing products | 3-5 years | 21:30 |
| Power reduction after switching to voltage-driven components | 1-2 orders of magnitude | 24:31 |
| Photonic neuromorphic lags electronic neuromorphic by | about 10 years | 37:32 |
| MIT photonic processor 6G signal classification latency | 120 nanoseconds | 43:33 |
| MIT photonic processor 6G classification speed advantage | 100 times faster than digital | 43:33 |
| Power savings of optoelectronic analog memory in the paper | 26 times | 42:33 |
Glossary
- WDM (Wavelength Division Multiplexing)
- A technique that sends multiple signals simultaneously through the same fiber using different colors of light, multiplying information throughput.
- Semiconductor Optical Amplifier (SOA)
- A device that amplifies optical signals directly without converting to electrical signals; in this scheme, it also performs linear and nonlinear operations.
- McCulloch-Pitts neuron model
- The most basic mathematical model of a neuron, consisting of a linear weighted sum followed by a nonlinear threshold activation.
- Indium Phosphide (InP)
- A semiconductor material platform that can integrate lasers, detectors, and optical amplifiers on a single chip, used for photonic integrated circuits.
- Femtojoule per operation
- A unit measuring energy consumption per operation; one femtojoule equals 10^-15 joules, the target energy efficiency range for photonic computing.
- TDM (Time Division Multiplexing)
- A method that transmits different signals over the same physical channel by allocating time slices; stacking with WDM can further expand communication capacity.
How to listen
Engineers and entrepreneurs concerned with AI compute bottlenecks, neuromorphic computing researchers, and investors wanting to assess whether photonic computing is worth betting on.