I'm reading some thing about CDR and how to implement clock recovery, more specifically the "Bang Bang" method. I quickly made a VCO (current starved) schematic and that can definitely provide a a frequency of my data source but the range of frequency it can do is fairly wide, including very far from the data frequency, so although I can see how the bang bang method can "keep track" of the alignement and tweak the VCO to stay in lock, I'm a bit more dubious about its ability to acquire a lock at the right frequency in the first place. Are you supposed to somehow first bring the VCO to close to the expected data rate "somehow" ? And have some secondary control voltage ?

Some PLLs have a coarse and a fine tune. The coarse tune is done in the circuit initialization. During operation, the PLL changes only the fine tuning, so the frequency range is more limited. This helps with tracking and jitter.

Yeah, I looked around and several paper seem to indeed point that problem that the VCO needs a large range (because of PVT variation you need to be able to tune it in all case), but that the typical Bang Bang will not be able to achieve lock starting from nowhere. I saw several approach, but the most applicable for me seem to indeed be use a reference with a simple FLL to achieve coarse lock and then switch to bang-bang for actual tracking. And since I also want to do that for USB HS where it's "packets", that meant basically locking to the reference until an active signal is detected at the input and then switch to tracking the signal.

I don't know much about the actual implementation of SerDes. The little I understand is that there is an initial code for the PLL to track in the beginning of a transmission, then comes the real packets of data. But there must be an initial phase for coarse tuning. I think there is already a reference signal for the receiver. Two PLLs. I'm a bit ashamed, because i work with that and I'm absolutely clueless about the big picture.

That's a really interesting paper, thanks. The sequential approach is what I was leaning toward.