Surface
Lolly Zheng- Sales Account Manager at NextPCB.com
NextPCB Capabilities
Printed Circuit Boards
PCB Assembly
Layer Buildup
SMD-Stencils
PCB Design-Aid & Layout
Mechanics
Quality
Drills & Throughplating
Factory & Certificate
PCB Assembly Factory Show
Certificate
Support Team
Feedback:
support@nextpcb.com
PCIe Gen6 (PCIe 6.0) doubles the per-lane throughput of Gen5 from 32 GT/s to 64 GT/s, delivering approximately 256 GB/s across a ×16 link—twice the bandwidth of Gen5. On paper, this looks like a straightforward generational improvement. In practice, Gen6 achieves the doubling not by increasing the signal frequency (which remains 32 GHz Nyquist, the same as Gen5) but by switching from two-level NRZ (Non-Return-to-Zero) signaling to four-level PAM4 (Pulse Amplitude Modulation). That change has profound consequences for PCB design that go far beyond simply tightening the existing Gen5 rules.
PAM4 encodes two bits per symbol by using four distinct voltage levels. At the same symbol rate as NRZ, PAM4 carries twice the data. But the spacing between voltage levels in PAM4 is one-third the spacing in NRZ at equal signal amplitude. This three-fold reduction in voltage margin means that noise, reflections, crosstalk, and jitter sources that were manageable in Gen5 NRZ become channel-killing problems in Gen6 PAM4—unless every element of the PCB design is tightened accordingly.
PCIe Gen6 is implemented in NVIDIA's B200 GPU (SXM6 form factor) as the host CPU interface, and is expected to become standard across data center AI hardware through 2026 and beyond. Layout engineers who have mastered Gen5 design will find that Gen6 requires a step change in rigor—not just incremental improvements.
Every PCIe generation from Gen1 through Gen5 used NRZ signaling: each symbol period carries one bit, encoded as either a high voltage level (logic 1) or a low voltage level (logic 0). Doubling data rate with NRZ requires doubling the symbol rate, which doubles the Nyquist frequency and doubles the frequency at which the channel must maintain adequate signal quality. By Gen5 (32 GT/s, 16 GHz Nyquist), the PCB channel loss budget was already fully utilized with premium low-loss laminates.
Doubling again to 64 GT/s with NRZ would require 32 GHz Nyquist performance from the PCB channel. At 32 GHz, dielectric and conductor losses are so severe that no commercially available laminate can sustain a channel of practical length (50–100 mm) within a usable loss budget. PCIe Gen6 instead uses PAM4, which keeps the symbol rate at 32 GT/s (16 GHz Nyquist—the same as Gen5) but encodes two bits per symbol by using four amplitude levels: 0, 1, 2, and 3.
The trade-off: PAM4 at the same amplitude swing as NRZ has three times smaller voltage spacing between adjacent levels (the eye height for each of the three eyes in a PAM4 eye diagram is one-third the NRZ eye height). Any noise, jitter, or distortion that would have been comfortably within the NRZ eye opening may now cause bit errors in PAM4.
The PCB design consequence is not a harder version of the same problem—it is a qualitatively different problem where every source of signal degradation (dielectric loss, conductor loss, via reflections, reference plane disruption, crosstalk) must be controlled to a standard that leaves the PAM4 eye open after all impairments are summed.
| Parameter | PCIe Gen5 | PCIe Gen6 |
|---|---|---|
| Line rate (GT/s) | 32 | 64 |
| Signaling | NRZ (2-level) | PAM4 (4-level) |
| Symbol rate | 32 GBaud | 32 GBaud |
| Nyquist frequency | 16 GHz | 16 GHz |
| Bits per symbol | 1 | 2 |
| Effective bandwidth per lane | ~4 GB/s | ~8 GB/s |
| ×16 link bandwidth (bidirectional) | ~128 GB/s | ~256 GB/s |
| Encoding | 1b/1b (effective) | 242b/256b FLIT |
| FLIT mode | No (optional) | Yes (mandatory) |
| FEC | Mandatory (CRC-based) | Mandatory (stronger; FBER-based) |
| Differential impedance | 85 Ω ± 10% | 85 Ω ± 10% |
| Max channel insertion loss | < 28 dB at 16 GHz | < 28 dB at 16 GHz (same limit, harder to meet with PAM4 margins) |
| Voltage margin per eye | Full swing | ⅓ of NRZ equivalent swing per eye level |
The identical Nyquist frequency between Gen5 and Gen6 (both 16 GHz) is a key point: the PCB channel insertion loss specification is nominally the same (< 28 dB at 16 GHz). However, the PAM4 system has one-third the voltage margin of an equivalent NRZ system. This means that channel impairments (reflections, crosstalk, jitter) that consumed say 10% of the Gen5 NRZ eye budget will consume 30% of the Gen6 PAM4 eye budget at the same absolute magnitude. In practice, Gen6 PCB channels must achieve significantly lower insertion loss, return loss, and crosstalk than Gen5 channels to leave adequate PAM4 eye opening after equalization.
PCIe Gen6 introduces FLIT (Flow Control Unit) mode as mandatory. A FLIT is a 256-byte packet that includes a CRC for error detection. The Gen6 FEC scheme adds redundancy within each FLIT to enable correction of burst errors that CRC alone would only detect. Two FEC modes are defined in the Gen6 specification:
The pre-FEC BER requirement (< 10−3) is dramatically more relaxed than Gen5's pre-FEC channel quality expectation, reflecting the reality that PAM4 channels inherently have higher raw error rates than NRZ channels at the same channel loss. Gen6 FEC is designed to clean up a PAM4 channel that looks poor in raw BER terms but whose errors are correctable with reasonable overhead.
Gen6 FEC adds more latency than Gen5 FEC because of the larger FLIT size and stronger correction capability:
For AI server workloads where PCIe is used primarily for model loading, checkpoint writing, and non-latency-critical host-to-GPU data transfers, this additional FEC latency is not a performance concern. For latency-sensitive inference applications where PCIe round-trip time directly affects time-to-first-token, the Gen6 FEC latency contribution should be included in system latency budgets.
FEC is designed to correct random and burst errors up to the pre-FEC BER limit. It does not fix:
These limitations underscore why FEC is not a substitute for careful PCB design—it is a safety net for residual random errors after a well-designed channel has already minimized systematic impairments.
The PCIe Gen6 specification nominally maintains the same 28 dB insertion loss limit at 16 GHz as Gen5. However, this parity is deceptive. The PAM4 system requires that the channel deliver adequate signal quality for the receiver's equalizer to separate four voltage levels, not two. The effective channel insertion loss that leaves a usable PAM4 eye after equalization is substantially lower than the specification limit:
PAM4 channels are more sensitive to return loss (impedance discontinuities) than NRZ channels because reflections cause inter-symbol interference (ISI) that affects all four voltage levels simultaneously, creating pattern-dependent eye closure that is difficult for equalizers to fully compensate. Gen6 return loss requirements:
| Loss Contributor | Gen5 Typical | Gen6 Target (practical) | Notes |
|---|---|---|---|
| PCB trace (dielectric) | 10–18 dB | 8–14 dB | Shorter channels or better material required |
| PCB trace (conductor / skin) | 3–8 dB | 2–5 dB | HVLP copper foil strongly recommended |
| Via structures | 1–4 dB | 0.5–2 dB | Tighter stub control; optimized anti-pad geometry |
| Connector loss | 0.5–3 dB | 0.5–2 dB | Gen6-rated connectors only; verify S-params at 16 GHz |
| Package launch | 1–2 dB | 0.5–1.5 dB | Optimized pad geometry; use vendor S-parameter model |
| Practical total target | < 28 dB (spec limit) | < 20–24 dB (design target for PAM4 margin) | Leave ≥ 4–8 dB margin vs spec limit for PAM4 eye adequacy |
The Gen6 differential impedance target remains 85 Ω, the same as Gen5. The specified tolerance is also ± 10%. However, the effective tolerance that keeps the PAM4 channel within margin is tighter in practice:
The Gen6 Nyquist frequency (16 GHz) is identical to Gen5. However, because Gen6 requires a smaller practical channel loss budget (20–24 dB design target vs < 28 dB spec limit) to maintain PAM4 eye margin, the effective laminate quality requirement is stricter:
Conductor loss at 16 GHz is significant for both Gen5 and Gen6, but the reduced PAM4 margin in Gen6 means that every dB saved on conductor loss directly translates to PAM4 eye margin:
| Laminate | Df (10 GHz) | Copper Foil | Gen6 Suitability | Max Practical Channel Length (Gen6) |
|---|---|---|---|---|
| Standard FR4 | ~0.020 | Standard ED | Not suitable | < 30 mm (unusable) |
| Panasonic Megtron 6 | ~0.004 | LP | Marginal; very short channels | < 80 mm with optimized vias |
| Isola Tachyon 100G | ~0.0021 | VLP | Good | Up to 150–180 mm |
| Panasonic Megtron 6E | ~0.0024 | VLP | Good | Up to 150–170 mm |
| Panasonic Megtron 7 | ~0.0020 | HVLP | Best practice | Up to 200 mm; used in B200 boards |
| Rogers RO4350B | ~0.0037 | VLP | Conditional (check budget) | Up to 120 mm with careful via design |
Via stub resonance creates insertion loss notches in the channel frequency response. For Gen5 NRZ, a notch at or above the Nyquist frequency (16 GHz) was acceptable if the notch depth was within the insertion loss budget. For Gen6 PAM4, even a partial notch whose skirt increases loss at 12–16 GHz by 2–3 dB can be sufficient to close the PAM4 eye.
The required minimum stub resonance frequency for Gen6 is higher than for Gen5 to keep the notch skirt effects below 12 GHz:
For Gen6 PCB channels, through-hole vias with residual stubs after backdrilling are acceptable only if the stub length is < 0.5 mm (approximately 20 mils). Achieving this with standard ± 50 μm backdrill depth accuracy requires very careful stackup design to ensure that the signal layer is sufficiently close to the drilled-from face to leave an acceptable stub.
For the most loss-sensitive Gen6 channels (maximum length, multiple layer transitions), laser-drilled blind/buried microvias that eliminate stubs entirely are the preferred solution. This is consistent with the HDI technology used on B200 boards for both NVLink 5.0 and PCIe Gen6 routing. For HDI design details, see Why AI GPUs Require 30+ Layer HDI PCBs.
For Gen6, via pad and anti-pad geometry optimization follows the same principles as Gen5 but with tighter targets:
PCIe Gen6 transmitters use Feed-Forward Equalization (FFE) with multiple pre-cursor and post-cursor taps to pre-distort the transmitted signal so that it arrives at the receiver with a more open eye after the channel's frequency-dependent attenuation. Gen6 TXEQ requirements:
The important PCB design implication: TXEQ can compensate for smooth, frequency-monotonic channel attenuation (the type caused by dielectric loss and conductor loss), but it is less effective at compensating for notches in the insertion loss response caused by via stub resonances or abrupt impedance discontinuities. A channel with a smooth 22 dB insertion loss may be more trainable than a channel with 18 dB average loss but a 3 dB notch at 14 GHz.
Gen6 receivers combine multiple equalization stages:
Equalization is more capable on channels with smooth frequency-dependent loss than on channels with resonances or mode conversion. PCB design decisions that affect equalizability:
The maximum practical PCIe Gen6 trace length (without a retimer) on a well-designed board is shorter than Gen5:
For AI server baseboards where the CPU-to-GPU distance may be 200–300 mm, Gen6 channels of this length almost always require a PCIe retimer. This is consistent with the design of B200 NVL72 and GB200 architectures where retimers are commonly used on PCIe Gen6 channels of moderate length.
Intra-pair skew requirements for Gen6 are tighter than Gen5:
The tighter requirement reflects the PAM4 system's greater sensitivity to any signal that converts differential mode to common mode. A 1 ps skew between P and N at 32 GBaud creates approximately −30 dB of differential-to-common mode conversion, which is at the limit of the Gen6 mode conversion budget.
Achieving < 1 ps intra-pair skew requires:
Gen6 PAM4 crosstalk requirements are functionally more stringent than Gen5 NRZ because the PAM4 eye has less absolute voltage margin to absorb crosstalk noise:
Reference plane disruptions are more damaging to Gen6 PAM4 than Gen5 NRZ because any impedance discontinuity or mode conversion caused by plane disruption reduces the already-tight PAM4 eye margin. Gen6 reference plane rules are stricter:
A PCIe retimer is a signal conditioning chip that receives, re-equalizes, retimes, and retransmits the PCIe signal partway through a long channel. It effectively splits a single long channel into two shorter channels, each within the loss budget independently.
For PCIe Gen6, retimers are more frequently required than for Gen5 because the effective maximum channel length is shorter due to PAM4 margin constraints. Key design considerations when using Gen6 retimers:
| Design Parameter | PCIe Gen5 (NRZ) | PCIe Gen6 (PAM4) |
|---|---|---|
| Line rate | 32 GT/s | 64 GT/s |
| Signaling | NRZ | PAM4 |
| Nyquist frequency | 16 GHz | 16 GHz (same) |
| Effective voltage margin per eye | Full NRZ swing | ⅓ NRZ swing per PAM4 eye |
| Insertion loss spec limit | < 28 dB at 16 GHz | < 28 dB at 16 GHz (same spec, tighter practical target) |
| Practical design target IL | < 26 dB | < 20–24 dB (need PAM4 eye margin) |
| Differential impedance | 85 Ω ± 10% | 85 Ω ± 10% (target ± 5% in practice) |
| Minimum laminate Df | ~0.003 (Megtron 6E) | ~0.002 (Megtron 7) |
| Copper foil minimum | LP (VLP recommended) | VLP minimum; HVLP best practice |
| Via stub target length | < 10 mils | < 5 mils; < 10 mils absolute max |
| Backdrill depth accuracy | ± 50 μm | ± 25–50 μm (tighter preferred) |
| Intra-pair skew | < 1.5 ps per segment | < 1.0 ps per segment |
| Lane spacing (edge-to-edge) | 3W rule minimum | 4W rule minimum |
| TX/RX layer separation | Strongly recommended (same layer possible) | Required (separate layers with ground plane between) |
| Reference plane gaps allowed | > 1 mm from traces | > 2 mm from traces |
| Max channel length (no retimer) | 200–250 mm | 150–200 mm |
| Retimer need | Optional for long channels | Commonly required > 150–200 mm |
| FEC type | CRC-based; pre-FEC BER < 10−4 | FBER-based; pre-FEC BER < 10−3 acceptable |
| Simulation requirement | Recommended for long channels | Required for all channels > 100 mm |
| Manufacturing complexity | High | Very high |
NVIDIA's B200 GPU (Blackwell, SXM6 form factor) uses PCIe Gen6 ×16 as its primary host CPU interface, providing approximately 256 GB/s of bidirectional CPU-to-GPU bandwidth. In the GB200 Superchip, the Grace CPU connects to the B200 GPU via NVLink-C2C (900 GB/s coherent interconnect) rather than PCIe, but external PCIe Gen6 connections to other system components (NIC, storage, management) are still present.
On B200 baseboards, PCIe Gen6 signal routing coexists with NVLink 5.0 in the same stackup. Since NVLink 5.0 already demands Megtron 7 laminate and HVLP copper on its routing layers, PCIe Gen6 benefits automatically from the board's premium material selection. The additional constraint is routing isolation: NVLink 5.0 differential pairs running at 200 Gb/s per lane generate electromagnetic fields that must be shielded from adjacent PCIe Gen6 traces by adequate spacing (≥ 6W between routing groups) or separate layer assignment. For a complete treatment of B200 board design constraints, see NVIDIA Blackwell Architecture Explained: B200, GB200 & PCB Design Impact.
Looking beyond B200, PCIe Gen6 is expected to become the standard AI server accelerator host interface through 2026–2027, with PCIe Gen7 (128 GT/s per lane, likely PAM4 again at higher symbol rate) appearing on roadmaps for 2028+. Layout engineers investing in Gen6 design mastery—especially the transition from NRZ to PAM4 thinking and the associated PCB design discipline—are building skills that will apply through at least the next two PCIe generations.
Pre-layout simulation is not optional for Gen6 channels longer than 100 mm. The reduced PAM4 eye margin makes it essential to verify that the channel meets requirements before committing to a board design. The recommended simulation workflow:
Why does PCIe Gen6 use PAM4 instead of simply doubling the symbol rate?
Doubling the Gen5 symbol rate to 64 GBaud would require 32 GHz Nyquist performance from the PCB channel. At 32 GHz, dielectric and conductor losses in any commercially available laminate are so high that even a 50 mm channel would consume the full insertion loss budget. PAM4 at 32 GBaud (16 GHz Nyquist—same as Gen5) achieves double the data rate by using four voltage levels rather than two, keeping the channel frequency requirements within range of achievable PCB designs—albeit with much reduced noise margin compared to NRZ.
Can a PCIe Gen5-rated PCB be used for Gen6?
It depends on the channel length and the specific materials used. A PCIe Gen5 board designed with Megtron 7 and HVLP copper, < 5 mil via stubs, and < 150 mm channels may meet the PCIe Gen6 practical channel requirements, particularly if a retimer is used for longer segments. A PCIe Gen5 board designed with Megtron 6 and LP copper and relying on the full 28 dB budget will not support Gen6 without redesign. The key question is whether the channel achieves < 20–24 dB insertion loss at 16 GHz with adequate PAM4 eye margin—not whether it meets the Gen5 specification.
Is backdrilling still required for PCIe Gen6 if I use HDI microvias?
No. Laser-drilled blind/buried microvias have no stubs by definition and do not require backdrilling. One of the key benefits of HDI microvia technology for Gen6 is that it eliminates the stub resonance problem entirely on the via transitions that use microvias. Through-hole vias on Gen6 signal layers still require aggressive backdrilling (< 5 mil stub target) if microvias cannot be used at that location.
What is FLIT mode and why is it mandatory in Gen6?
FLIT (Flow Control Unit) mode restructures PCIe transactions into fixed 256-byte packets (FLITs) that carry their own CRC and are managed as atomic units. FLIT mode is mandatory in Gen6 because it enables the Gen6 FEC mechanism to operate on well-defined packet boundaries, and because the 242b/256b encoding used in FLIT mode is more efficient than the per-byte overhead of earlier PCIe encoding schemes. Gen5 introduced optional FLIT mode; Gen6 makes it non-optional.
How many retimers does a B200 baseboard typically need for PCIe Gen6?
It depends on the board layout. In a DGX B200-class system where the host CPU and GPU packages are on the same baseboard with PCIe Gen6 channels of 150–250 mm, one retimer per ×16 link is typically required. A system with 8 B200 GPUs and 2 CPUs could require 8 or more PCIe Gen6 retimers, each consuming 1–3 W of power and requiring PCB area for placement and routing. Retimer count is a meaningful cost and complexity driver in B200 baseboard design.
What is the pre-FEC BER target for a PCIe Gen6 channel?
The PCIe Gen6 FEC mechanism is designed to correct burst errors when the pre-FEC BER is < 10−3. In practice, a well-designed channel should achieve significantly better than this: pre-FEC BER of 10−6 to 10−5 is a reasonable design target that leaves adequate headroom above the FEC correction limit for process variation and aging. A channel operating routinely near the 10−3 FEC correction limit is marginal and may degrade over the product lifetime as connector and board aging increase channel loss.
PCIe Gen6 PAM4 boards demand the tightest combination of laminate quality, copper foil selection, via stub control, and impedance precision available in commercial PCB manufacturing. NextPCB supports Megtron 7 and ultra-low-loss laminate processing, HVLP copper, ± 5% impedance control, precision backdrilling to < 5 mil stubs, HDI microvia technology, and complete PCBA services for AI server and data center PCB programs.
Related Articles:
Still, need help? Contact Us: support@nextpcb.com
Need a PCB or PCBA quote? Quote now
