--- title: "PAM4 vs. NRZ Modulation in Transceivers: The Practical Implications" slug: "pam4-vs-nrz-modulation-transceivers" type: deep-dive category: "Modulation & Signal Integrity" tags: [PAM4, NRZ, modulation, CDR, link-budget, 400G, fiber-quality, signal-integrity] seo_focus_keyword: "PAM4 vs NRZ transceiver modulation" --- The migration from NRZ (Non-Return-to-Zero) to PAM4 (Pulse Amplitude Modulation, 4-level) is the defining signal engineering change of the 100G-to-400G transition. Every engineer deploying 400G or higher-speed optics needs to understand what PAM4 actually means for their fiber plant, their link budget assumptions, and what happens when the two technologies coexist in the same infrastructure. ## NRZ: The Baseline NRZ is the modulation scheme that has dominated optical networking since digital transmission began. In NRZ, each bit period contains a single signal level representing either a 1 or a 0. A 100G NRZ system uses 4 parallel lanes, each running at 25 Gbps NRZ — the baud rate (symbol rate) equals the bit rate per lane. This is conceptually clean: one baud = one bit. The noise sensitivity of NRZ is determined by the eye opening — the vertical and horizontal separation between the 0 and 1 signal levels at the decision point. A clean NRZ eye has roughly half the optical power range available for separating 0 from 1. Optical sensitivity for a 25G NRZ receiver is typically around -10 to -14 dBm minimum for a good transceiver, depending on the specific implementation. 25G NRZ at the lane level has been successfully deployed on OM3 and OM4 multimode fiber for distances from 1 m to 300 m (OM4), and on OS2 single-mode fiber for distances up to 10 km (10GBASE-LR profile scaled to 25G lane rates). The technology is proven, the fiber requirements are well-characterized, and the installed base is enormous. ## PAM4: Four Levels Instead of Two PAM4 encodes 2 bits per symbol by using four distinct amplitude levels instead of two. A 50 Gbaud PAM4 signal carries 100 Gbps per lane. This is why 400G QSFP-DD can use 8 lanes at 50 Gbaud PAM4 rather than requiring 16 lanes at 25 Gbaud NRZ — the lane count stays manageable as bit rates scale. The fundamental tradeoff: PAM4 squeezes four amplitude levels into the same voltage or optical power range that NRZ used for two. The three eye openings in a PAM4 signal (between levels 0-1, 1-2, and 2-3) are each only one-third the size of the single NRZ eye opening. This means the noise margin for each level transition is significantly reduced. Specifically, for the same peak optical power, PAM4 has 9.5 dB less margin per eye compared to NRZ. That 9.5 dB is why PAM4 systems require much stronger FEC (Forward Error Correction). 100G NRZ over SR4 uses KR4 FEC or KP4 FEC as optional. 400G PAM4 over SR8 or DR4 mandates KP4 FEC — without FEC, the raw BER from a PAM4 system running at the edge of its link budget would be unacceptable. KP4 FEC can correct a raw BER of up to 2.4×10^-4 to below 10^-15, which is what makes PAM4 practical despite the noise margin reduction. ## The Link Budget Difference A 100G SR4 link (4 lanes, 25G NRZ per lane) over OM4 has an application code loss budget of 1.9 dB for the channel loss at 850 nm (this covers the fiber, connectors, and splices, not the transceiver internal loss). The minimum launch power per lane is typically -4 dBm and the receiver sensitivity is around -9.5 dBm, leaving 5.5 dB of component margin beyond the fiber channel budget. A 400G SR8 link (8 lanes, 50G PAM4 per lane) over OM4 has an application code channel loss budget of 1.9 dB as well — the same number. But the minimum launch power per lane is now 0 dBm (versus -4 dBm) and the receiver sensitivity is -6.5 dBm for a good implementation. The effective operating range is similar on paper, but the PAM4 system is running the lasers harder and requiring better receiver performance to achieve the same optical channel. The practical consequence: a fiber infrastructure that "works fine" for 100G SR4 may be marginal for 400G SR8, even at the same distances, if: The connectors are at or near the 0.35 dB per mating specification rather than the 0.10–0.15 dB typical for good connectors. On a 100G NRZ link, 5 connector pairs at 0.30 dB each (1.5 dB total) leaves 0.4 dB of headroom. On a 400G PAM4 link where the PAM4 eye margin reduction has already consumed most of the engineering margin, the same 1.5 dB of connector loss may push the link outside the acceptable operating region, especially on temperature variations. The fiber itself has higher attenuation than nominal. OM4 is specified at 3.5 dB/km at 850 nm. Old OM4 that has been routed through tight bends, patched dozens of times, or is running warm in poorly ventilated trays may measure 4.0–4.5 dB/km at 850 nm. For a 50 m run, that's still under 0.25 dB extra — not significant. For a 100 m run at the edge of spec, it matters. ## CDR Requirements and What Happens Without Them Clock Data Recovery is the DSP function that synchronizes the receiver sampling clock to the incoming signal. In NRZ systems at 25G, CDR is helpful but not always mandatory — many short-reach multimode links run without CDR in the receiver because the eye opening is large enough for a simple comparator. This is why some 10G and 25G SFP28 short-reach modules are sold as "CDR-free" variants, which are cheaper and have lower latency. PAM4 systems at 50G per lane require CDR in both the transmitter and receiver. The transmitter CDR is needed because the ASIC serializer outputs a NRZ signal from two NRZ lanes, which the CDR converts to PAM4 before the optical interface. The receiver CDR performs the inverse: PAM4 optical to two-lane NRZ electrical. There is no "CDR-free" PAM4 transceiver for 50G+ lane rates because the DSP is integral to the modulation scheme. The implication: PAM4 transceivers have higher power consumption (more DSP), more latency (CDR adds roughly 20–50 ns), and more points of failure (the DSP itself) compared to equivalent-bandwidth NRZ systems. This is a known and accepted tradeoff at 400G and above. ## When Your Old Fiber Plant Fails The most common deployment failure pattern with PAM4 is this: an operator deploys 400G DR4 or FR4 transceivers on existing single-mode fiber infrastructure that was installed for 10G or 40G. The fiber tests clean at 1310 nm with an OTDR. The links come up. A few weeks later, some links are flapping or showing elevated FEC counters. The fiber plant may be fine in terms of attenuation. The problem is chromatic dispersion accumulation. PAM4 at 50G per lane is more sensitive to dispersion than 10G NRZ because the symbol period is shorter (20 ps for 50 Gbaud vs. 100 ps for 10G NRZ) and the 4-level eye opening is smaller — dispersion-induced pulse spreading closes an already-small PAM4 eye faster than it closes an NRZ eye. 400G DR4 has a dispersion tolerance of approximately ±50 ps/nm, which translates to about 310 m of SMF-28 at 1310 nm. For runs under 500 m, dispersion is not the issue. For runs of 1–2 km on older SMF that runs slightly higher dispersion per km, it can be. The practical takeaway is that "the fiber worked before" is not sufficient qualification for PAM4. Test the actual insertion loss at the operating wavelength, count the connector matings in the path, and check the fiber type specification. For 400G and above, the fiber infrastructure needs to meet a tighter tolerance than it did at 100G NRZ, even if the nominal link budget numbers look similar.