Shock headline, lots of caveats. April 2013 release of QLogic Windows device driver for 2[4-6]xx Fibre Channel HBAs (22.214.171.124) uses 64-bit DMA addressing. Presumably previous versions (<=126.96.36.199) use 32-bit DMA addressing. Supporting evidence from Release Notes:
Added support support for full 64 bit physical addressing [ER101742]
Have not talked with anyone from QLogic to confirm this, but seeing the above in the Release Notes was an a-ha! moment. I believe it because I have lots of observations that fit. Specifically using 100% of a core per 200 MB/s I/O stream (probably doing buffer copying and interrupt handling) and extreme volatility of I/O speeds under load. Now I know why Emulex cards perform better in Windows 2008 R2 environments.