The gm of the inverter was my first concern. But I tuned the bias current by +/-25%, it did not change anything. The good chip is still good, bad is still bad. The gm is 3.9/4.9/5.6m correspondingly(by simulation). Humungus, do you have an idea what the typical gm should be?
I also suspected the output current is too high. Because from the spec of the crystal, the drive leve is 100uW Max. Converting to current, it's 1.6mA. But the RMS of the output current by simulation is 2.7mA. But the crystal didn't fail, as I said, I once swapped the crystal to check, the good chip is still good, the bad is still bad. I am really confused.
The process variation shouldn't cause so big difference, should it?
Sorry for the late reply.