"Saturating your links" is rarely the goal in HFT.

You want low deterministic latency with sharp tails.

If all you care about is throughput then deep pipelines + lots of threads will get you there at the cost of latency.