Impeller will create min(4, hardware_concurrency) threads to boot strap VkPipelines after creating the context_vk. On the Mokey deivce, using 4 threads is actually much slower than using 1. The Mokey is only a 4 core device so at a minimum we should reduce the max concurrency to something like hardware_concurrency / 2.

4 thread (~3 seconds)

1 thread ( ~ 2 seconds)

We should figure out how to adjust the concurrency level based on device performance.