Lesson 20 of 25 · 15 MIN · 80% complete

High-Speed Transceivers

GTX/GTH/GTY transceivers are dedicated serial I/O blocks that enable PCIe, 10GbE, Aurora, JESD204B, and other multi-gigabit protocols — completely separate from general IO banks, with their own PLLs, CDR, equalization, and encoding engines.

What are GTX/GTH/GTY Transceivers?

Transceivers are dedicated high-speed serial I/O blocks that exist completely outside the general IO bank structure. They handle protocols at line rates from 600 Mbps to 58 Gbps per lane — far beyond what regular LVDS IO can achieve. Each transceiver is self-contained with its own PLL, serializer, deserializer, CDR, encoding, and equalization:

Protocols supported: PCIe Gen1/2/3, 10 Gigabit Ethernet, SATA, SAS, DisplayPort, JESD204B, Aurora, USB 3.0
Not available in all devices — present in Kintex-7, Virtex-7, Zynq-7000 (partial); Artix-7 has GTP (lower rate)
Organized in quads: 4 channels per quad, each quad sharing one QPLL and individual CPLLs

Transceiver Family Comparison

Type	Max Line Rate	Typical Use	Device Family
GTP	6.25 Gbps	USB 3.0, PCIe Gen2, SATA	Artix-7 (some SKUs)
GTX	12.5 Gbps	PCIe Gen3, 10GbE, Aurora	Kintex-7, Virtex-7
GTH	16.375 Gbps	40GbE, OTN, backplane	Virtex-7, Zynq UltraScale+
GTY	32.75 Gbps	100GbE, CFP	UltraScale+
GTM	58 Gbps	400GbE, PAM-4	Versal

Serial Link Architecture

Each transceiver channel is full-duplex with completely independent TX and RX paths. The TX serializes parallel data from the FPGA fabric and drives a differential output; the RX receives the incoming differential signal and recovers both clock and data before deserializing back to parallel:

flowchart LR subgraph TX["TX Path"] direction LR TXDATA([Parallel\nData]) --> ENC8B["8b/10b or\n64b/66b\nEncoder"] ENC8B --> SER["Serializer\nGEARBOX"] SER --> PREEMPH["Pre-emphasis\nDriver"] PREEMPH --> TXPAD(["TX+/TX−\n(pad)"]) end subgraph RX["RX Path"] direction LR RXPAD(["RX+/RX−\n(pad)"]) --> EQ["Equalizer\nLPM / DFE"] EQ --> CDR["CDR\nPLL lock\nto data"] CDR --> DESER["Deserializer\nGEARBOX"] DESER --> DEC8B["Decoder\n8b/10b or\n64b/66b"] DEC8B --> RXDATA([Parallel\nData]) end style ENC8B fill:#1e3a5f,stroke:#3b82f6,color:#93c5fd style CDR fill:#1e3a5f,stroke:#3b82f6,color:#93c5fd style SER fill:#1e293b,stroke:#475569,color:#94a3b8 style DESER fill:#1e293b,stroke:#475569,color:#94a3b8

Clock Recovery — CDR (Clock and Data Recovery)

The receiver has no separate clock input from the transmitter. Instead, it uses a CDR (Clock and Data Recovery) circuit to extract the transmitter's clock from the transitions in the received data stream:

The CDR PLL locks its phase and frequency to detected data transitions
Lock time: typically 1–10 ms depending on data pattern and PLL bandwidth
RXCDRLOCK status bit — the design must check this bit before trusting recovered data
8b/10b encoding guarantees sufficient transitions for CDR lock; raw binary data without coding can cause CDR loss-of-lock

PLLs in Transceivers: QPLL and CPLL

Each GTX quad contains two types of PLL for generating the high-rate serial clocks:

QPLL (Quad PLL): one per quad, shared across all 4 channels — used for high line rates (>6.6 Gbps). Lower jitter than CPLL at high rates.
CPLL (Channel PLL): one per channel — used for lower rates, allows each channel to run independently at a different rate
PCIe Gen3 (8 Gbps): use QPLL — all lanes share the same reference clock and PLL
Aurora at 3.125 Gbps: CPLL is adequate and allows mixed-rate quads

QPLL sharing constraint Each GTX quad has 4 channels and 1 QPLL. The QPLL is shared — all 4 channels must run at the same or harmonically related line rates. For designs mixing PCIe and SATA in the same quad, use CPLL per channel or assign different protocols to different quads.

Equalization and Pre-emphasis

Long PCB traces, connectors, and cables attenuate high-frequency content, causing inter-symbol interference (ISI) that closes the received eye diagram. Transceivers compensate using:

TX Pre-emphasis: the transmitter intentionally boosts high-frequency content before driving the line. Configured via TXPRECURSOR, TXMAIN, TXPOSTCURSOR registers (each in 1/64 units of total voltage swing)
RX Linear Equalizer (LPM): a linear filter that boosts high frequencies at the receiver — suitable for short to medium channels (<10 dB loss)
RX DFE (Decision Feedback Equalizer): an adaptive equalizer that uses past symbol decisions to cancel ISI — required for channels with >10 dB insertion loss

Equalization Mode	Best For	Power
LPM (Linear)	Short channels, FR4 PCB traces <15cm	Low
DFE (Adaptive)	Long channels, backplanes, cables	Higher

8b/10b and 64b/66b Encoding

Serial links require sufficient signal transitions for CDR lock and DC balance. Xilinx transceivers support two encoding schemes natively in hardware:

Encoding	Data bits → Line bits	Bandwidth overhead	Used in
8b/10b	8 → 10	20% overhead	PCIe Gen1/2, SATA, Aurora, JESD204B
64b/66b	64 → 66	3.1% overhead	PCIe Gen3, 10GbE, 40GbE, 100GbE

8b/10b replaces each 8-bit data byte with a 10-bit code word chosen to guarantee at least 3 transitions per 10 bits and DC balance (equal 1s and 0s). The 20% bandwidth penalty is the price of this guarantee.

Protocol Reference

Protocol	Line Rate	Encoding	Typical GTX Config
PCIe Gen1	2.5 Gbps/lane	8b/10b	CPLL or QPLL, x1/x4/x8 lanes
PCIe Gen2	5.0 Gbps/lane	8b/10b	QPLL preferred
PCIe Gen3	8.0 Gbps/lane	64b/66b	QPLL required
10GbE (XFI)	10.3125 Gbps	64b/66b	GTX QPLL, 1 lane
Aurora 8b/10b	3.125 Gbps/lane	8b/10b	CPLL, 1–4 lanes
JESD204B	3.072–12.5 Gbps	8b/10b	QPLL for >6.6G

Xilinx Transceiver Wizard IP

Manually configuring a GTX requires setting over 100 attributes correctly. Xilinx provides the Transceiver Wizard IP to automate this entirely:

Open Vivado IP Catalog → Transceiver Wizard (7 Series)
Select transceiver type (GTX, GTH, GTP), line rate, reference clock frequency, and protocol
The wizard calculates all QPLL/CPLL M/N divider values and sets all GTX attributes
Output: complete Verilog wrapper, simulation model, XDC constraints with REFCLK pin assignments
Includes an example design with a self-test loopback mode for hardware bring-up

Always use the Transceiver Wizard GTX has 100+ configuration attributes — setting them manually is error-prone and not supported by Xilinx for general use. The Wizard generates validated configurations with simulation models and example designs. Never manually configure GTX attributes from scratch.

Eye Diagram and Signal Quality

An eye diagram is the primary tool for evaluating received signal quality at multi-gigabit rates. It overlays many successive bit periods to show the statistical distribution of voltage and timing:

Open eye: wide vertical opening → good signal integrity, low BER
Closed eye: traces collapse to the center → excessive ISI, reflections, or noise → bit errors
Measured using Vivado IBERT (Integrated Bit Error Ratio Tester) — a debug core inserted into the design that sweeps voltage/time margins and plots a bathtub curve
IBERT communicates via JTAG while the transceiver operates at full line rate

GTX reference clock pin requirement GTX reference clocks must come from dedicated REFCLK pins (MGTREFCLK0 / MGTREFCLK1 per quad), not from BUFG or general IO. Using a non-dedicated reference clock source will cause the QPLL to fail to lock and Vivado will report a critical warning.

Interview question Q: What is CDR in a transceiver and why is it needed?
A: CDR (Clock and Data Recovery) extracts the transmitter's clock from the received data stream. Since the receiver has no separate clock input from the transmitter, CDR uses a PLL that locks onto data transitions to reconstruct the correct sampling clock. Without CDR, the receiver cannot sample the incoming data reliably.

Knowledge Check

Q1. How many channels are in one GTX quad?

Correct! Each GTX quad contains exactly 4 independent transceiver channels, all sharing one QPLL. This is why PCIe x4 and x8 configurations must stay within one or two quads respectively.

Q2. What does CDR stand for?

A Channel Data Rate
B Clock and Data Recovery
C Calibrated Delay Reference
D Continuous Data Routing

Correct! CDR — Clock and Data Recovery — is the circuit that extracts the transmitter's clock from the received serial data stream by locking a PLL to the data transitions.

Q3. Which GTX PLL is shared across all 4 channels in a quad?

A CPLL (Channel PLL)
B QPLL (Quad PLL)
C MMCM
D PLL_ADV

Correct! The QPLL (Quad PLL) is shared across all 4 channels in a quad. It is preferred for high line rates (>6.6 Gbps) due to lower jitter. The CPLL is a per-channel PLL used for lower rates or mixed-rate quads.

Q4. PCIe Gen3 uses which encoding scheme?

A 8b/10b
B 64b/66b
C NRZ with no encoding
D 128b/132b

Correct! PCIe Gen3 switched from 8b/10b to 64b/66b encoding to reduce the bandwidth overhead from 20% down to 3.1%. This is why Gen3 at 8 Gbps raw line rate delivers ~7.877 Gbps effective throughput per lane.

Q5. What Xilinx debug core measures bit error rate on transceivers at full line rate?

A ILA (Integrated Logic Analyzer)
B VIO (Virtual IO)
C IBERT (Integrated Bit Error Ratio Tester)
D ChipScope Pro AURORA

Correct! IBERT is a dedicated Vivado debug IP that inserts alongside the transceiver to sweep TX swing, pre-emphasis, and RX equalization settings while counting bit errors and generating eye diagram / bathtub curve data via JTAG.

Q6. GTX reference clocks must come from which source?

A Any BUFG-driven clock in the design
B MMCM output divided to the required frequency
C Dedicated REFCLK pins (MGTREFCLK0 / MGTREFCLK1)
D Any general IO pin with LVDS standard

Correct! GTX reference clocks must use the dedicated MGTREFCLK0 or MGTREFCLK1 pins of each quad. These feed directly into the QPLL/CPLL with the low-jitter path required for multi-gigabit locking. Using general IO or BUFG will cause QPLL lock failure.

Practical Exercise

Use the Vivado Transceiver Wizard (IP Catalog → 7 Series FPGAs Transceivers Wizard) to generate a GTX configuration with the following requirements:

Protocol: Aurora 8b/10b
Number of lanes: 4 (full quad)
Line rate per lane: 3.125 Gbps
Reference clock: 156.25 MHz (from MGTREFCLK0)
Encoding: 8b/10b (hardware)

After generation, note:

Which PLL was selected — CPLL or QPLL? What are the M and N divider values?
What is the TXOUTCLK frequency delivered to the FPGA fabric?
What are the recommended TXPRECURSOR and TXPOSTCURSOR pre-emphasis values?

Open the generated example design and review the top-level port list. Identify the REFCLK, TXP/TXN, RXP/RXN, and GTXRESET ports. Simulate the example design using the provided simulation model to observe TXOUTCLK and RXRECCLKOUT timing.

← IO Banks Config Memory →