What's the perf/$ ratio for these different instances ?

I've also read that the A100 can deliver 4x the performance on some DL workloads.