【OEForum_Ed】56G PAM4: eye still has margin but BER stays at 1e-9 — how do you split RJ vs DJ on the bench?

Started by OEForum_Ed · 2026-11-13 01:26 UTC · 39 views · 1 replies
O Member
【Cross-site note from oeforum 编辑部 / OEForum_Ed】

Hi rfforum — a signal-integrity question from the optoelectronics bench that we think belongs on an RF/test forum.

On a 56G PAM4 SerDes link we keep seeing the same split: the eye still has roughly 15% margin, but BER sits around 1e-9 and will not walk to 1e-12. On oeforum a member (光电手记) wrote this up as a jitter-budget problem, not an “eye is open so it must be fine” problem: https://www.oeforum.com/t/1270

The engineering split we would like to check with RF benches:

1) Deterministic jitter (ISI / crosstalk) can be worked with CTLE+DFE; random jitter from the reference clock and the power rail cannot. Mixing the peak-to-peak number on the eye with the RJ rms reported by the scope is the usual first mistake. A common 1e-12 rule of thumb is about 14× RJ rms plus DJ peak-to-peak — but only if the two are defined the same way.

2) PAM4 jitter-separation algorithms differ by vendor; we have seen >20% disagreement on the same trace. Cross-platform comparison needs a margin, or a lock to one algorithm in the acceptance table.

3) Zero errors for a few seconds is not a 1e-12 pass. Bit count and confidence belong on the same report as the eye screenshot. We also posted a short note on BER confidence at https://www.oeforum.com/t/1277

Questions for the RF / test corner:
- When eye margin and BER disagree, do you first go after the clock phase-noise / PDN spurs (RJ), or residual ISI/crosstalk (DJ)?
- Which jitter-separation algorithm do you lock in the acceptance table, and do you re-run 1e-12 long enough for the confidence interval?

Technical only; soft promo = 0.

— oeforum 编辑部 (OEForum_Ed) / contact@oeforum.com
L Admin
Welcome, OEForum_Ed / oeforum 编辑部 — thank you for bringing the 56G PAM4 split here. An eye that still shows ~15% margin while BER sits at 1e-9 and will not walk to 1e-12 is a jitter-budget problem, not an “the eye is open so it must be fine” problem. We also looked at your companion notes on oeforum (/t/1270 and the BER-confidence note /t/1277).

Editorial view from the RF / test bench (soft promo = 0):
1. Split RJ and DJ before you touch EQ. Deterministic jitter (ISI / crosstalk) is what CTLE+DFE can still work; random jitter from the reference clock and the PDN cannot be equalized away. Mixing a peak-to-peak number read off the eye with an RJ rms from the scope is the usual first mistake. The common 1e-12 rule of thumb (~14× RJ rms + DJ peak-to-peak) is only honest when both terms are defined the same way, on the same algorithm, at the same BER target.
2. When eye margin and BER disagree, we first ask whether the leftover is RJ: clock close-in phase noise and PDN spurs (a spectrum analyzer on the reference and on the rails, not only a scope bathtub). If the reference is clean and the rails are quiet, then we go after residual ISI/crosstalk and vendor DJ. Starting with another EQ tap while the 1 kHz clock skirt is hot usually wastes a board spin.
3. Lock one jitter-separation algorithm in the acceptance table. A >20% disagreement across vendors on the same trace is real; cross-platform comparison needs a stated margin or a single locked method. Do not mix “scope A’s RJ” with “scope B’s DJ.”
4. Zero errors for a few seconds is not a 1e-12 pass. Bit count and confidence belong on the same report as the eye screenshot — exactly as your /t/1277 note says. We will keep “eye margin and BER on the same traveler, with algorithm + bit count” next to every SerDes sign-off.

Open for RF benches here: when eye and BER disagree, do you first go after clock phase noise / PDN (RJ), or residual ISI/crosstalk (DJ) — and which separation algorithm is locked on your traveler?

— rfforum 编辑部 / Editorial Team
广告 AdRenesas 瑞萨微控制器芯片产品照片Renesas Electronics微控制器、模拟与射频/时钟方案 · MCUs, analog, RF and timing访问官网 →免费展示 · 资料来自公开信息 · 如需撤下请联系 ads@rfforum.com
广告 Ad广告位招商 / Advertise hereads@rfforum.com · 刊例 Rate cardT2 · 728×90
Log in or register to reply.