Playground · Photonic accuracy demo
How much noise can a photonic tile take?
The photonic tensor accelerator (PTA) is in design, so it is measured against an error model before anything is built.
Here that model runs the canary, a small MNIST classifier whose last layer is on a 128 × 64 photonic tile, on 1,000 test images.
Change the noise and see whether it still passes Gate F.
Real harness The canary harness from kubeworkz/grxcp, compiled to WebAssembly and running in your browser. Its CI publishes a build only when it prints exactly what the native harness prints. This one: 98954ac.
The canary's settings:
The first run downloads about 2.8 MB.
Accuracy
Gate F · error
Gate F · class agreement
Converter saturations
What the harness printed, and how to run it yourself
Natively, from docs/designs/canary/harness built with -DN_SIDE=8 -DN_HID=64 -DN_OUT=16:
Gate F, as ruled for the canary (manifest.json): the normalised RMS error with one output scale must fall in a physical band of 2.5% to 7.3%, and the scaled outputs must agree with the reference's class on at least 98.5% of the images. Error below the band is better than a physical device is expected to manage. About the PTA →