I am running Samsung M393B2G70BH0-YK0 (1600 MHz ECC 16 GB sticks validated at 1.35v @1600) at 1866 MHz and 1.55v, prime stable. I could run it @2133 MHz, too without changing of the timings but after several hours of heavy-duty use system throws machine check exception, meaning unrecoverable ECC error has occurred. I suppose loosening timings and bumping the voltage would help but I decided not to bother since I moved my algorithms to CUDA anyway.
Problem with 1866 MHz might not be due to memory alone - it could also be due to CPU's own IMC. Although Intel's processes are superb nowdays and most CPUs end up pretty fine when it comes to tolerances, it is still possible that your Xeons simply do not allow 1866 MHz DDR3 speed without bumping the VSa voltage.
You can try to bump VSa a bit. As for the DDR3 voltage, personally I would not go higher than 1.55v but for the sake of hypothesis validation I guess running a test with, say, 1.60v would not matter too much.
I am using Noctua NH-D14 (the big ones) and the clearance between the heatsink and PCIe x16 slot #1 is too small for a fan. However, I do not care too much at the moment since I am not able to use GTX TITAN when connected to PCIe slot #1 anyway (I strongly suspect PCIe link training failures).
Yeah, I haven't tried bumping up the voltage at all. I thought I was getting another kit of RAM to run in this board, but that fell through, so I may do some more playing around with the kit I have.
It probably doesn't hurt that you are running ECC too.