I guess these NICs have a round PCI clock signal so DMA fails often and overwrites kernel memory with zeros, especially during high loads.
Anyway, we're going to replace them with good cards.
We could switch everything to use mod_backhand again during the downtime, but that's enough work that's it's not worth it. If I had tools already written to switch all the settings I'd do it, but I don't.
Anyway, the site will be down in about an hour for maybe 5 minutes. I'd do it at night if I could, but this time is best for the guy helping me.