The description for writeTimestamp currently states:

Writes a timestamp value into a querySet when all previous commands have completed executing.

Notably, this doesn't guarantee that subsequent commands haven't started yet. As a result, code like this:

writeTimestamp(A)
do_some_gpu_work(...)
writeTimestamp(B)

can actually be reordered by the GPU to be:

do_some_gpu_work(...)
writeTimestamp(A)
writeTimestamp(B)

resulting in the measured time being unreasonably short.

(This is related to, but not exactly the same as gfx-rs/wgpu#2521)