We know that we need to make some key architectural changes to wgpu's resource management, but wgpu's performance is a key selling point, and it's difficult to assess the impact of the changes we're considering without having benchmarks at hand.

The ideal would be something in CI, but we're sort of at zero right now, and we need some way to answer engineering questions. The benchmarks could be informal, require manual steps, be microbenchmarks, require use of external tools like Tracy - all would be a step forward.