One of the coolest parts of this is 'pipelining' which allows us to schedule the prove tasks while the executor is still running. So with ideal saturation of GPUs to execute speed our e2e proving time is only a few % less than execution speed.
(Scaling this workflow is a fun job)
We've had the ability to run arbitrarily large programs in ZK and verify on-chain for a quite a while now, but I still think how it works is cool.
