pgbench and scaling

PG 9.5.3
RHEL Linux

To cut the story short and giving as much details as I can remember:

I created 40 database db1 .. db40 in a cluster.
On each database I initialized pgbench tables with a scale of 15.

Then I concurrently ran 5 pgbenches for 15 min. Each pgbench connected to one db. In other words, no two pgbench
sessionswere working on the same db. 

I noticed that as I scale from 5 to 10 to 20 to 40, the TPS starts falling almost linearly :

with 5, TPS was doing 639
with 10 TPS was down to 490
with 20 TPS was down to 280
and so on.

Are we missing something?  Where do I start looking for the root cause.

thanks