Interoperability of bucket values

[RByers]

I think implementation-defined bucket values is probably right (there's likely no good way to formally define a precise specific). That said, I also imagine that rough interoperability for them will prove to be important. As soon as websites start altering the experience based on the performance bucket, it will become important to site developers that they get a broadly similar experience across browsers on similar devices. And then it will become important to device makers to try to predict what performance class they will get assigned for new devices to maximize the user experience.

This is perhaps similar to the debate we're currently having of interoperability of on-device AI APIs which are inherently probabilistic. I think we're converging on the use of industry benchmarks as a helpful but not overly constraining tool. I'd suggest something similar here. Maybe we should publish a sample site that computes some value based on signals our implementation uses? Or perhaps we should publish a database of popular devices and their value (all non-normative of course)?

One other tool which may help is for Chromium to have a user setting for overriding the predicted value, similar to how most games allow the user to choose a higher-than-recommended graphics quality setting for their device (trading eg. resolution for framerate). Knowing the user can alter the value may also help with the privacy debate.

[handellm]

I like the idea of a game-like override UI. In the presence of features gated on the performance tier, it becomes handy for example if the built-in prediction algorithm mispredicts.