There's ongoing work in #4711 and #4842 to improve benchmarks, but getting the data requires looking at the build artifacts, which are also not persistent (I seem to remember buildkite logs are deleted after 6 months or so, but don't quote me on that).

For benchmarks run on main we can collect the results somewhere (like a branch in this repo) and display them in a wesbite, something like Enzyme-JAX or Reactant.jl benchmarks, which are both powered by https://github.com/benchmark-action/github-action-benchmark. We can also explore other visualisation tools.