I've observed the same behavior on AMD and Intel at $WORK. Our solution (ideal for us, reads happening roughly 1B times more often than writes) was to pessimize

I've observed the same behavior on AMD and Intel at $WORK. Our solution (ideal for us, reads happening roughly 1B times more often than writes) was to pessimize writes in favour of reads and add some per-thread state to prevent cache line sharing. We also tossed in an A/B system, so reads aren't delayed even while writes are happening; they just get stale data (also fine for our purposes).