Best way to run GPU containers

Bringing existing containerized solution into AWS. I have few questions. I uploaded containers to ECR, created RDS etc. created cluster and task definitions, started, all good. Now need to figure out how to run containerized custom nvidia triton in AWS. This needs 1-2 A100/80GB or similar.

1. What is the best instance to run one or two A100 GPUs. If instance is p4d.x24large can i use only one GPU from it or does billing run for all 8?
2. Planning to turn GPU task on only when needed, for example 8 hours a day and destroy when not needed. I investigated and there are many ways to do it but what is the recommended way?
3. Should i use service discovery to get the GPU server ip and port info as this would be in task2 while container that accesses this is in task1 - is there simpler method for the case when GPU server is started- destroyed/stopped. If instance is started and destroyed the system will give it new ip adderesses. Is there better way. The idea is to save $$$ in GPU cost.

Thanks in advance!