QPI Laboratories, Inc. — Photonics · Simulation · Inference
QPI Laboratories builds a simulator for photonic AI accelerators — where you find the compute sweet spot before it becomes silicon.
QPI Labs provides the first publicly accessible simulator that runs real neural-network models on a photonic-compute abstraction. Developers can explore photonic matrix-vector multiplication, multi-wavelength compute, and photonic inference behavior inside full AI model graphs — long before hardware exists.
Manifesto
Every extra order of magnitude in model size now costs more in electrons than in math. Electronic compute burns energy on every operation and stalls on every clock cycle, while the math keeps outgrowing the machines that run it.
Light computes as it propagates — free of clocks and heat. Freeform approaches go further still: beams interfere in open space, in whatever geometry the optics allows. Interference does the linear algebra while light simply travels.
QPI Laboratories, Inc. builds a simulator for photonic AI accelerators — the tool that finds the compute sweet spot before committing to hardware. The hardware and the roadmap stay behind the curtain for now; today we are building the place where you feel the physics first.
Status
The simulator is live. The full picture stays in the dark — where light behaves best.
// The Simulator — Live
Emitter array 28×28 — MNIST 7
Photonic pipeline
VCSEL Array
narrow-line · coherent
MZM Weight Matrix W
thin-film LiNbO₃
Photodetector
Σᵢ xᵢ·T(wᵢⱼ)
Readout
Model: MNIST · Linear
Photonic accuracy: 92.6%
Ideal accuracy: 98.1%
Class probabilities
Run real neural-network inference on a photonic-compute abstraction — tune noise, quantization, and hardware parameters in real time.
Launch SimulatorWho it's for
Five teams, one simulator — each arrives with a different question and leaves with the same answer.
Your architecture is still on the bench. The simulator lets you pressure-test modulator choices, non-ideality budgets, and array sizing in software — before a single chip is fabbed.
You're building the chip that runs the model. The simulator tells you what the model costs in photonic hardware — and where the bottleneck will be before your first tapeout.
Quantization, pruning, and mixed-precision decisions all have a photonic cost. Run your compressed model through the simulator before assuming the hardware can absorb the trade-off.
— Grounded in the Literature
The simulator's physics models, non-ideality equations, and throughput analysis are derived from the following landmark papers.
Deep learning with coherent nanophotonic circuits
Shen et al. · 2017
Inference in artificial intelligence with deep optics and photonics
Wetzstein et al. · 2020
Parallel convolutional processing using an integrated photonic tensor core
Feldmann et al. · 2021
Large-scale optical neural networks based on photoelectric multiplication
Hamerly et al. · 2019
Fundamentals of Coherent Optical Fiber Communications
Kikuchi · 2016
— Who We Are
QPI Laboratories was founded in 2026 by Dr. Andreas Falkenberg — serial inventor at the intersection of photonics and machine learning. The mission is singular: make photonic AI compute explorable before it becomes silicon.
We believe the transition from electronic to photonic inference is inevitable. The researchers and engineers who lead that transition need simulation infrastructure precise enough to inform real hardware decisions — before a single chip is fabbed. The QPI Labs simulator is the first step.
What the simulator does
The QPI Labs simulator runs real neural-network models on a photonic-compute abstraction. It lets developers explore photonic matrix-vector multiplication, multi-wavelength compute, and photonic inference behavior inside full AI model graphs — long before hardware exists.
Every inference is computed as a physical matrix-vector multiply — optical signals through weighted modulators, summed by photodetectors.
WDM parallelism assigns independent wavelength channels to separate compute paths, multiplying effective throughput without scaling array size.
Run full pretrained MLPs and CNNs through the photonic pipeline — seven models from MNIST Linear to CIFAR-10 ResNet-8 are included.
— What's inside the simulator
— Physics Reference
A full photonic compute textbook. Baked in.
21 sections of rendered physics — LaTeX transfer functions, intuition callouts, and wins/losses tables for every modulator and noise source. Every equation the simulator uses is documented and explained.
— Live Comparison
Ideal and photonic inference run side by side.
Every run shows IDEAL FLOAT32 and PHOTONIC WITH NOISE simultaneously — including full confusion matrices for both. The gap between them is the hardware cost, made visible.
— Hardware Sizing
Eight live hardware sliders. One bottleneck chart.
Tune array width, modulation rate, ADC/DAC sample rates, weight precision, and batch size — throughput and bottleneck recompute in real time. The per-tile time budget stacked bar shows exactly where your cycles go.
— Design Space Sweep
Sweep any parameter. Plot the Pareto frontier.
Pick a non-ideality parameter, enter comma-separated values, add an optional second axis for a 2D heatmap, run — the simulator plots accuracy vs throughput across every configuration. Hardware co-design at simulation speed.
— Dual Physics Paths
Incoherent or coherent. Two completely different physics.
Switch to coherent detection and the system uses homodyne readout — signed weights emerge naturally from the interference term. A full mode change with its own physics engine.
— Download
Run a sweep. Get publication-ready figures.
After any Pareto sweep, hit Download — the simulator generates a ZIP of clean, formatted plots ready for a paper or slide deck. The figures arrive ready to use.
Everything runs on real NumPy compute — real physics, real numbers.
The Simulator
Our web-based simulator finds the compute sweet spot before you commit to hardware: swap between five modulator technologies, choose your laser source and detection mode, inject real non-idealities, and sweep the Pareto pages for accuracy against cost — all in your browser.
// access is provisioned by a QPI Labs administrator on request
Open the Simulator— Request Access
Accounts are provisioned manually. Drop your email and a line about your work — we'll reach out.
