I understand that pricing is usage based on Railway. How fine grained is this measurement? Do you round up at all during calculations?

I ask because: I want to setup horizontal autoscaling for my application, which is not natively supported by Railway, and I am limited to providing a fixed number of replicas.

But let's say I have 10 replicas. Each is configured to 1CPU and 1GB RAM.
But average usage turns out to be 0.1CPU and 0.1GB RAM per replica.

Will I then pay a total for 1CPU + 1GB RAM?

Or do you round up the usage from each replica before summing? (and if you do round up, then to what extent? nearest 1CPU? 0.1CPU?)

In other words, I could have 50 replicas. My traffic per replica would then go down 50x vs 1 replica. If I do this, would my "idle usage overhead" be the only additional cost I need to pay for?