"send: No buffer space available" on EL9 when using Zero-copy

I found a strange inconsistency between running neper with -Z on a Alma 9 and Ubuntu 22.04 hosts, both using default kernel.
With default configs both works fine (mtu 1500 and max tcp_wmem 4MB)
But when MSS is above 3444 and max tcp_wmem above 9MB, on Alma 9 "send: No buffer space available" keeps spamming on the client side. On Ubuntu 22.04 it works with no errors.

Example config that works on both OSes: MTU=1500 and net.ipv4.tcp_wmem="4096 16384 4194304"
Example config that gives errors on Alma 9: MTU=4168 and net.ipv4.tcp_wmem="4096 16384 16777216"

My test environment are 2 servers back to back, with the same spec and same Connectx-6 NIC. I'm just dual booting between both OSes to tshoot.
I looked both servers sysctl and NIC configs using ethtool and couldn't find any differences.