We have a design using two DP83867E Phys (connected together with the internal switch of a T1020 CPU via SGMII). Generally the design works very well. However, there are a few specific problematic cases that we have observed, possibly with the same or related root causes. With one link partner our device consistently takes several times longer to negotiate a link and the link doesn't always seem to work properly. With another link partner, the pair polarity and MDIX mode resulting from auto-negotiation varies significantly and the link often doesn't work even after it claims auto-negotiation completed successfully. We had our device analyzed using a PVA-3000 PhyView®Analyzer from Sifos and the resulting report showed that the TX amplitude in Gigabit mode was over 2db below nominal on all pairs, at about 1.2V P-P. I'm not sure if this is related to the symptoms we are seeing, but it seems plausible. Both of the ports on our device use a different Jack, both with integrated magnetics, but since both ports behave the same even with these different Jack/magnetics parts, I don't believe that is the issue. For reference, one part is Bel 0821-1C1T-43, and the other is Wurth 7499111221A.
I was hoping to find a way to tune the transmission power or amplitude of the DP83867E but I didn't see anything in any documentation about that. I'm also looking for any other information on diagnosing problematic links. I have compared a register dump from working and non-functional links and so far nothing has consistently stood out (besides the variations in auto-negotiation results cited above) that suggests a cause. I have also verified the reference resistor value, the voltage across this resistor during operation, and the power supply voltages, and all are nominal.