An Under-the-Hood View of PCIe 3.0 Link Training (Part I)

Figure 1: An overview of the elements of PCIe 3.0
dynamic link equalization

Now that we've looked at the basics of PCIe 3.0 dynamic link equalization and at some of the particulars of de-emphasis and preshoot, it's time to dive a little deeper into what actually happens in the link training process. It all happens in the blink of an eye but there's enough going on to warrant some dissection.


On the transmit side of the channel is a three-tap finite impulse response (FIR) filter with three settable cursor coefficients; these cursors determine the level of de-emphasis and preshoot before the signal enters the channel. As discussed in our last post, presets may be applied to these settings. On the receive side, a behavioral equalization algorithm comprises continuous time linear equalization (CTLE), decision feedback equalization (DFE), and clock/data recovery (CDR). This algorithm may equalize the signal after it has traversed the channel (Figure 1).

Figure 2: Main state diagram for an LTSSM

All of the above must take place at runtime because when we plug a PCIe add-in card into a host system, the designers of those respective systems couldn't know which cards would be added to which systems. Thus, the channel parameters are unknown, so the selection of TxEQ cursors or presets, the tuning of CTLE and DFE for RxEQ all must be configured at runtime so the add-in card and system can work out the best equalization settings to maintain an acceptable error rate at the 8-GT/s data rate. And this is where dynamic link equalization enters the picture.

In the PCI-SIG's language, two PCIe devices exchange "training sequences" to negotiate a number of link parameters, including elements such as lane polarity, link/lane numbers, equalization, data rate, and so on. The way this happens is through the execution of a link training and status state machine (LTSSM), which is depicted in Figure 2. When a PCIe add-in card is plugged into a system and everything is powered up, the LTSSM will be implemented in the order of the blue-circled states. The Recovery state includes a number of substates shown in Figure 3, one of which is recovery.equalization (outlined in green). That is where all the link training takes place to adjust the equalization parameters for that particular channel.

Figure 3: The Recovery substates

Inside of recovery.equalization are four stages of the link tuning and equalization, labeled Phases 0 through 3 (Figure 4), enabling both sides of the link to advertise their capabilities and get into the training process. A starting-point preset is the entry to further tuning of the link, so the system provides that to the add-in card while still at 2.5 GT/s.

Phase 0 starts at a rate of 2.5 GT/s. The system (upstream) sends a speed-change request along with a TxEQ preset to the add-in card (downstream). That preset is to be used when the add-in card first jumps to 8 GT/s. In this fashion, the system can establish communication with a bit-error rate of at least 10^-4

Figure 4: A diagram of the four phases of
recovery.equalization

Next, the add-in card moves to 8 GT/s using the TxEQ preset sent by the system. In Phase 1, the system and add-in card advertise their equalization capabilities to each other. In Phase 2, the downstream add-in card adjusts the upstream system's TxEQ settings while tweaking its own RxEQ settings. In Phase 3, that process reverses; the system adjusts the add-in card's TxEQ while setting its own RxEQ.

When all of the above is accomplished and we exit out of recovery.equalization, we should be at the optimal TxEQ and RxEQ on both sides of the link and should be operating at a BER of at least 10^-12.

In the next installment of this series of posts on PCIe 3.0 dynamic link equalization, we'll take a closer look at the meat of the process, which occurs in Phases 2 and 3 of recovery.equalization.