Compile to
Light.
Route neural network inference through silicon photonic chips. Matrix multiplications at the speed of a laser pulse — not a clock cycle. 0.4ns latency. Real hardware. No simulation.
Photon vs GPU vs TPU.
No asterisks.
All benchmarks run on identical workloads: PyTorch models compiled via photon.compile(), measured end-to-end including I/O. Hardware: Photon SPC-7 card, A100 80GB, TPUv4.
All measurements: median of 10,000 runs · batch_size=1 · fp16 precision · Feb 2026
Every layer of the stack.
Fully specified.
Compiler Toolchain
PyTorch → photonic IR → silicon
Photon's frontend accepts any PyTorch nn.Module or JAX function. A single decorator triggers compilation to photonic IR without modifying model architecture.
import photon@photon.compile(target="spc-7", precision="fp16")def model_forward(x):return resnet50(x)# Compiles on first call, cached to .photon/cache/output = model_forward(input_tensor)
The optimizer maps tensor operations to physical MZI arrays, minimizes optical path length, and resolves phase conflicts across wavelength channels.
Photonic computation's energy cost scales with precision. The quantization engine finds the optimal precision floor per layer using calibration data.
photon.quantize(model,calibration_data=loader,target_precision="opt4", # optical 4-bitper_layer=True)
One command.
Inference at light speed.
No email form. No waitlist. Paste the command, run your model, measure the nanoseconds yourself.
Browser Sandbox requires GitHub OAuth — no email, no credit card. Your models never leave your browser session.
Engineers who measured it
themselves.
"I've benchmarked every accelerator on the market. Photon is the first time I've had to rewrite my measurement code because my timer resolution wasn't fine enough. 0.4ns is not a marketing number."
"Our Series B investors asked us to prove the photonic accelerator claim. I ran photon bench, exported the signed PDF, and sent it over. Due diligence closed in three days. That report is worth more than the hardware."
"The compiler correctly maps our custom attention variant to MZI arrays without manual waveguide routing. I expected to spend three weeks on the HAL integration. It was a Thursday afternoon."
