Asus AC58U Performance regression with 21.02

After upgrading to 21.02 from 19.07 I have noticed that the WAN speed went from ~900mbps to ~750mbps. I have downgraded and done some benchmarks to confirm that this is indeed the case, here are the results: (note that I was testing it behind another router so the speedtest results have L2NAT, though the outer router's performance doesn't seem to be the bottleneck here. iperf3 only goes through AC58U's NAT). The normal LAN speed is ~940mbps.

19.07:

No software flow offloading:

speedtest: 838 down 842 up
iperf3: 814

With software flow offloading:

speedtest: 818 down 828 up
iperf3: 933

21.02 (with settings migration via LUCI):

No software flow offloading:

speedtest: 516 down 594 up (something wrong? 25% sirq cpu usage)
iperf3: 521

With software flow offloading:

speedtest: 760 down 782 up
iperf3: 682

Was something changed in 21.02 that could've caused this? The latest snapshots also have similar performance to 21.02.

Maybe it's the same problem as the AVM 4040 ...

Installing irqbalance does seem to improve things, however it's still about the same performance as 19.07 with software flow offloading enabled in both, and slower than 19.07 when software flow offloading is disabled.