The habitat of hardware bugs
A DRAM controller bug can be very damaging to the chip maker, whose engineering samples might not work and whose production schedule might be delayed. For the DRAM controller vendor - no big deal, "we have 3 more customers affected by this bug, we must say you're taking it unusually passionately!" This is an actual quote.
A chip can have 2 processors with 8-byte buses each, going to a DRAM giving you 16 bytes per cycle, through a shared 8-byte-per-cycle bottleneck. This interconnect is the handiwork of some time-starved dude on the chip maker's team, armed with an interconnect-generating tool. Even such an idiotic issue will manifest on some benchmarks but not others, and might not get caught at design time. And if you think that is stupid, I've heard of a level 2 cache which never actually cached anything, and this fact happily went unnoticed for a few months. (Of course, this not being caught at design time is when the team should start looking for a new career.)

https://www.embeddedrelated.com/showarticle/988.php