Programmable 200 GOPS Hopfield-inspired photonic Ising machine
Article Date: 17 December 2025
URL: https://www.nature.com/articles/s41586-025-09838-7
Summary
This Nature paper presents a room-temperature optoelectronic-oscillator (OEO) photonic Ising machine inspired by Hopfield networks. The architecture combines cascaded thin-film lithium niobate (TFLN) modulators, a semiconductor optical amplifier (SOA) and a digital signal processing (DSP) engine in a recurrent, time-encoded loop to implement programmable spin couplings and nonlinearity at very high throughput — reported at more than 200 GOPS. The system supports fully connected problems up to 256 spins (65,536 couplings) and scales to >41,000 spins for sparse graphs.
Key Points
- Room-temperature OEO-based photonic Ising machine inspired by Hopfield networks with linear spin representation scaling.
- Hardware uses cascaded TFLN modulators, a bulk SOA and an embedded DSP engine in a recurrent time-encoded loop.
- Demonstrated coupling and nonlinear operations at >200 GOPS, enabled by high baud-rate operation and DSP.
- Supports fully connected problems up to 256 spins and sparse graphs exceeding 41,000 spins; best-in-class photonic results on max-cut for 2,000 and 20,000 spins.
- Achieved ground-state solutions for benchmark optimisation tasks including number partitioning and lattice protein folding previously not solved by photonic systems.
- Uses intrinsic optical noise from high-baud-rate SOA operation as a useful mechanism to escape local minima and speed convergence.
- Experimental code and resources are available: https://github.com/Shastri-Lab/tfln-ising-nature-paper-2025.
Content summary
The authors build a programmable photonic Ising solver by time-multiplexing spins in an optoelectronic recurrent loop. Electrical signals from high-speed arbitrary waveform generators encode spins and weights; TFLN modulators imprint those onto light, while a bulk SOA provides gain and controlled noise injection. A DSP stack at both transmitter and receiver stages implements weight multiplication, readout and algorithmic control, improving convergence and solution quality compared with purely optical implementations.
Benchmarks presented include max-cut on arbitrary graph topologies (demonstrating strong performance at 2,000 and 20,000 nodes among photonic machines), number partitioning and lattice protein folding, with comparisons showing this platform attains competitive or superior solution quality in its class. The paper emphasises that embedding DSP techniques from optical communications into optical computation unlocks faster, more reliable optimisation runs.
Context and relevance
Ising machines are a hot approach for tackling combinatorial optimisation problems. This work links high-speed photonics, mature DSP methods and TFLN device performance to push practical, programmable photonic optimisation further. It addresses core bottlenecks: reconfigurability, speed and noise management, and demonstrates a route to scale beyond purely optical spatial approaches by exploiting time multiplexing and DSP.
For researchers and engineers in photonics, neuromorphic hardware and optimisation, the paper is a useful demonstration of how communications-grade hardware (TFLN modulators, SOAs, high-sample DAC/ADC and DSP) can be repurposed to deliver substantial GOPS-level compute for analogue optimisation tasks.
Author’s take (punchy)
This is a proper milestone for photonic optimisation — high-speed modulators + DSP = a seriously programmable Ising machine that plays well with real-world benchmarks. If you care about scaling optical optimisation away from lab demos and toward useful, reconfigurable hardware, read the methods and experiments in detail.
Why should I read this?
Short version: if you’re into fast hardware for NP problems (or just curious how optics + comms DSP can beat local search tricks), this paper shows a working, room-temperature system that actually solves non-trivial benchmarks. We’ve skimmed the heavy physics and pulled out the practical bits — read it if you want the how-to and the numbers without wading through mountains of equations.