Seek help understanding the profile trace of a GPU jit execution

[jjyyxx]

I collected profile trace with the following snippet

        with jax.profiler.trace("/tmp/xyz", create_perfetto_link=True):
            self.state, info, _ = stateless_update(key, self.state, data_tree_1, data_tree_2)

where stateless_update is a complex jitted function. I try to use profile trace to understand the most time-consuming part of the code.

However, it can be seen that there are many small gaps between GPU kernels in the GPU stream, and the sum of them (~4.7ms) is, of course, much smaller than the total function calling time (~11.5ms). Since I expect that most computation should take place on GPU, what am I missing here? Am I misinterpreting the results?