I just had to return my third instance of this motherboard. I'm going with a similar Supermicro, but with an LSI controller. All three boards have exhibited the same problems.
I have 6x2.5" Samsung 2TB and 2x2.5" Hitachi 1.5TB drives set up as a stripe of mirrors (3x2TB and 1x1.5TB on the 4 port Marvell controller make up one side of the mirrors, and 2x2TB on the Intel SATA3 ports, and the remaining two drives on the 2 port Marvell controller making up the other side of the mirrors). A Samsung 2.5" 2TB drive sits on one SATA2 port as a spare.
The controllers are all in AHCI and JBOD mode.
I have 4x8GB Crucial approved memory.
The first board died during a memory burn. After about a day, the system started a constant reset cycle. The BMI errors were showing a CPU overtemp, the system would shut down, and the CPU overtemp would IMMEDIATELY clear. The heat-sink was always cool to the touch.
Back it went to Amazon for a replacement.
The second board survived the memory test, and I started exercising the drives in sequence (a few errors) and then I started moving a data set over (4.5TB).
The minute I start a heavy write to the pool (rsync of 4.5TB of data from an older FreeNAS) the drives on the Marvell controllers start throwing CRC errors, with one or more drives occasionally disconnecting. A scrub on the full set (I did get the 4.5TB data over) will almost immediately disconnect one or more drives. The Intel connected drives never throw an error.
I contacted a very nice tech at Asrockrack, and he said the temp on the board and CPU should stay below ~50C. I was running up around 74C at load. I added more fans and ran them full-out (their SMART-FAN settings do not get it below 74C). That did not help.
I also swapped cables with a new set, and moved the drives around. The errors followed the Marvell ports, not the drives.
I had a further call with the Asrock folks, and they offered to swap the board, and burn-in a board at their facility before sending it to me.
I just got that board, and it shows the same errors - I'm done here. It's back in the box, heading for Amazon.
One thing on their testing - they tested a RAID 10 set on the 4 port Marvell controller, and a RAID 0 on the 2 port Marvell controller. That means that the parallel writes that ZFS does went through a different data path on the controllers (JBOD) vs. their test (which was one stream to the controller, which then spread the actions out to the drives). That might be the difference.