The current query incurs an expensive round-trip to the GPU thread, which can negate the benefits. To solve it we'd want a mechanism similar to how we check queries for completion (including polling from the GPU thread).