VDM and FEC metrics: PAM4 links

At 50G per lane and above, optical power stops being a sufficient health indicator: a PAM4 link with perfect Rx power can be one dirty connector away from uncorrectable errors, and the first sign is not a dBm but a bit error ratio. Two sources give that number. Inside the module, CMIS VDM (Versatile Diagnostics Monitoring) reports pre-FEC BER, SNR and laser health per lane; on the host, the FEC decoder counts corrected and uncorrected codewords. Together they show how far a link is from failure while it still carries traffic without loss.

Why FEC changed diagnostics

Every 25G+ Ethernet PMD on PAM4 (and most 25G NRZ ones) assumes Reed-Solomon FEC. The optics are allowed to deliver a raw error ratio of 10⁻⁵–10⁻⁴ because FEC repairs it to 10⁻¹² or better. The consequence: a degrading link shows no errors at all to the MAC until the FEC limit is crossed, then fails abruptly. The pre-FEC BER is the only smooth indicator of margin.

FEC	Code	Corrects per codeword	Pre-FEC BER limit for 10⁻¹² output (approx.)	Used by
BASE-R ("FC-FEC", Clause 74)	Fire code (2112, 2080)	11-bit bursts	≈ 10⁻⁷ … 10⁻⁸ gain of ~2 dB	10G/25G NRZ, CA-25G-S
RS-FEC KR4 (Clause 91/108)	RS(528, 514) over GF(2¹⁰)	7 symbols	≈ 5 × 10⁻⁵	25GBASE-CR/SR/LR, 100GBASE-SR4/CWDM4/CR4
RS-FEC KP4 (Clause 119/134)	RS(544, 514)	15 symbols	≈ 2.4 × 10⁻⁴	all 50G PAM4 lanes: 50G/100G/200G/400G/800G
Concatenated / oFEC (coherent)	inner + outer codes	—	≈ 10⁻² … 2 × 10⁻²	400ZR (CFEC), OpenZR+ (oFEC)

Configuration mismatches: FEC, auto-negotiation & link training.

VDM — what the module reports

VDM lives in CMIS pages 20h–2Fh (banked per 8 lanes) and is advertised in page 01h. Each observable has a descriptor (type, lane), a live sample, four thresholds and flags, exactly like the classic monitors.

Pages	Contents
20h–23h	descriptors: which observable, which lane, which threshold set
24h–27h	samples (2 bytes each)
28h–2Bh	thresholds: high/low alarm, high/low warning
2Ch–2Fh	flags and masks

Observable types (CMIS 4.0/5.x)

ID	Observable	Unit / format	What it tells you
1	Laser age	% (U16)	vendor's estimate of consumed life
2	TEC current	% of max (S16)	cooler load — Temperature & voltage
3	Laser frequency error	MHz (S16)	DWDM/ZR: distance from the grid
4	Laser temperature	°C (S16, 1/256)	cooled lasers
5 / 6	eSNR, media / host input	dB (U16, 1/256)	electrical SNR after equalisation — the PAM4 "eye height"; > 18–20 dB comfortable, < 15 dB marginal (vendor-dependent)
7 / 8	PAM4 level transition parameter, media / host	dB	linearity of the four levels
9 / 10	Pre-FEC BER minimum, media / host	F16	best over the interval
11 / 12	Pre-FEC BER maximum	F16	worst over the interval — the number to alert on
13 / 14	Pre-FEC BER average	F16	—
15 / 16	Pre-FEC BER current	F16	last sample
17–24	FERC (frame error ratio count) min / max / avg / current, media / host	F16	post-FEC uncorrectable frames — should be 0
128+	coherent set (CMIS 5.x): modulator bias, CD, DGD, SOPMD, OSNR, eSNR, Q-factor, Rx total/signal power, carrier frequency offset…	various	400ZR/ZR+ modules

F16 is CMIS's 16-bit ratio format: a 5-bit exponent and 11-bit mantissa, value = mantissa × 10^(exponent − 24). A pre-FEC BER of 1.5 × 10⁻⁶ is stored as mantissa 1500 with exponent 15. Media side = the optical line; host side = the electrical interface to the switch ASIC — a bad host-side BER points at the connector or the board, not the fibre.

Host-side FEC counters

The switch or NIC counts, per port and often per lane:

Counter	Healthy	Watch	Failing
Corrected codewords	growing slowly (pre-FEC BER 10⁻⁸ … 10⁻⁶)	fast growth (10⁻⁵ at KP4)	near the limit
Uncorrectable codewords	0	any increment	steady increments = frame loss
Symbol errors per lane	balanced across lanes	one lane dominating	that lane's fibre/connector/laser
Pre-FEC BER (derived)	< 10⁻⁷	10⁻⁶ … 10⁻⁵	> 10⁻⁴ (KP4)

Where to read them: show interfaces … fec / phy detail on most NOS (Verifying optics), ethtool -S (fec_corrected_blocks, fec_uncorrectable_blocks) and ethtool --show-fec on Linux, show interfaces counters fec on SONiC. Counters are cumulative — clear them and watch the rate.

Reading the two together

Rx power	Pre-FEC BER / eSNR	Uncorrectable	Diagnosis
in window	< 10⁻⁷, eSNR high	0	healthy with margin
in window	10⁻⁵, eSNR falling on one lane	0	that lane is degrading: dirty MPO fibre, ageing VCSEL, marginal connector — fix before it fails (Per-lane diagnostics)
low	high	some	budget problem — Rx power & link budget
in window, both ends	high on host side	some	electrical: cage/connector contact, board, DAC quality, wrong equalisation settings
in window	high on media side, all lanes	some	dispersion, wrong reach class, wavelength off the mux, polarity partially wrong
fine	0 errors, link flaps	0	not a signal problem — FEC/AN mismatch, thermal, software

Thresholds and alerting

VDM thresholds are set by the vendor; typical high-warning for pre-FEC BER is around a tenth of the FEC limit (2 × 10⁻⁵ at KP4). For alerting, use rates and trends: a pre-FEC BER climbing one decade in a week is a work order regardless of thresholds (Monitoring). Post-FEC uncorrectable errors are always an incident.

Where VDM is missing

SFF-8472 and SFF-8636 modules have no VDM: 10G/25G/100G NRZ links rely on host FEC counters and on Rx power. Some SFF-8636 modules expose vendor-specific SNR or BER in pages 20h–21h or vendor bytes — readable only with the vendor's tool. For NRZ without FEC (10G LR, 1G), CRC/input-error counters on the interface are the only bit-level view.

In CodingBox

For CMIS modules CodingBox reads the VDM descriptors and samples and shows them per lane alongside the classic monitors on DDM, decoding F16 ratios into plain numbers. Host FEC counters are outside the module and outside the programmer — they come from the switch.