Xian Yang Tee

Special expertise · High-speed test

High-speed test: bandwidth, eye diagrams and bit error ratio.

How fast a device is, what its data looks like, and how often it gets a bit wrong. This page sets out how I think about the three, what I built hands-on, and what I specified, kept clearly apart.

Hands-onOn-wafer RF bandwidth of photodiodes, integrated into an automated prober SpecifiedInstrument plan for 224 Gb/s PAM4: eye diagram and BER WherePhotonic wafer test, 2023 – now

Part 1 of 4

My approach

Bandwidth, eye diagram and BER answer different questions about the same signal.

1.1 · Frequency, time and statistics

How bandwidth, eye and BER fit together.

Three views of one high-speed link Left: frequency response S21 falling with frequency, with the 3 dB bandwidth marked. Middle: an eye diagram of overlaid bits, with eye height and eye width marked. Right: a bathtub curve of bit error ratio against sampling time, with the error floor at 1e-12 setting the usable timing margin. 1 · Frequency: bandwidth −3 dB f₃dB FREQ →S21 (dB) VNA / LIGHTWAVE ANALYSER 2 · Time: eye diagram EYE HEIGHT EYE WIDTH · JITTER · ISI PATTERN GEN + SAMPLING SCOPE 3 · Statistics: BER BER 10⁻¹² MARGIN SAMPLE TIME →log BER BERT: PRBS + ERROR DETECTOR LIMITED BANDWIDTH → INTER-SYMBOL INTERFERENCE → A CLOSING EYE → BIT ERRORS
Three views of the same link. Bandwidth in the frequency domain sets how cleanly symbols settle; the eye shows it in time; the BER counts what finally goes wrong.

A high-speed photonic or electrical link can be judged three ways, and each answers a different question.

  • Bandwidth (frequency domain). A swept S21 measurement on a vector network analyser, or a lightwave component analyser for optical-to-electrical and electrical-to-optical parts. The 3 dB bandwidth tells you whether a photodiode, modulator or receiver is fast enough before any data goes through it.
  • Eye diagram (time domain). A pattern generator drives the link and a sampling scope overlays thousands of bits. Eye height, width, jitter and inter-symbol interference show how close each decision is to failing.
  • Bit error ratio (statistics). A BERT sends a long pseudo-random pattern such as PRBS31 and counts errors. It is the final answer, and the slowest: proving 10⁻¹² takes on the order of 10¹² bits.
  • NRZ2 levels · 1 bit per symbol
  • PAM44 levels · 2 bits per symbol
  • 224 Gb/s PAM4112 GBd per lane
  • Nyquist frequencyhalf the baud rate: 56 GHz
  • PAM4 cost3 eyes, each ~⅓ the height: needs more SNR
  • Sampling scopelower noise than real-time, suits PAM4 eyes

Part 2 of 4

The problems I solved

Labelled hands-on or specified, so it is clear what I built and what I planned.

2.1 · Problem by problem

From on-wafer RF test to a 224 Gb/s instrument plan.

  1. 01Measuring a photodiode’s RF bandwidth before the wafer is diced
  2. 02Taking the test set out of the answer
  3. 03Getting optical fibres and an RF probe onto the same device
  4. 04Aligning the fibres when the signal is RF
  5. 05Making RF test part of the automated flow
  6. 06Choosing instruments for 224 Gb/s PAM4 testing
  7. 07Knowing how much receiver bandwidth a signal needs
  8. 08Testing receivers with built-in electronics
  9. 09A low reading: tester or device?

Bandwidth: measuring how fast a device is, on the wafer

01Industry · hands-on

Measuring a photodiode’s RF bandwidth before the wafer is diced

Why it’s hard
Bandwidth is normally measured on packaged parts; on the wafer the light arrives through edge couplers in trenches and the signal leaves through a probe.
What I did
Integrated a vector network analyser with an electrical-to-optical converter into the wafer prober: light from the converter enters the device through the fibre array, an RF probe picks up its response, and the analyser sweeps S21.
Result
S21 and 3 dB bandwidth measured at the wafer level, at several bias voltages.
02Industry · hands-on

Taking the test set out of the answer

Why it’s hard
Cables, probes and the converter all roll off with frequency; uncorrected, they look like a slow device.
What I did
Set up the calibration chain with my team: a 2-port RF calibration of cables and probes on a calibration substrate, then characterising the electrical-to-optical converter against a reference optical-to-electrical module, so its response is de-embedded from every device measurement.
Result
Bandwidth that belongs to the device, not the setup.
03Industry · hands-on

Getting optical fibres and an RF probe onto the same device

Why it’s hard
Fibre array and RF probe have to reach the same small device without colliding, and the pads must be placed for the probe.
What I did
Defined the probe configuration for RF test on photonic wafers, wrote design-for-test rules for where pads and couplers should sit, and developed revisions of the RF probe holder.
Result
Optical and RF probes landed together under the microscope.
04Industry · hands-on

Aligning the fibres when the signal is RF

Why it’s hard
Fibre alignment normally peaks on optical power; for an RF test the useful feedback is electrical.
What I did
Extended the alignment algorithm to take the photocurrent or the S21 value as its feedback.
Result
Fast alignment directly on the quantity being tested.
05Industry · hands-on

Making RF test part of the automated flow

Why it’s hard
A bench measurement does not scale to every die on a wafer.
What I did
Integrated the analyser through SCPI instrument drivers and RF test sequences in the control software, paired with the specific devices on the wafer map.
Result
RF bandwidth runs as one more automated step in the wafer test.

Eye diagram and BER: what a data link actually does

06Industry · specified

Choosing instruments for 224 Gb/s PAM4 testing

Why it’s hard
At 112 GBd the instruments are expensive and the trade-offs between them are not obvious.
What I did
Worked through the options with Keysight and Anritsu and wrote the recommendation: an arbitrary waveform generator for short research patterns versus a BERT pattern generator for long PRBS31 patterns; a lightwave analyser at 70–110 GHz for bandwidth; a sampling scope at 60 GHz for low-noise PAM4 eyes versus a real-time scope; and a lower-cost production BERT for volume.
Result
An instrument plan per test level: wafer, die, package and module.
07Industry · specified

Knowing how much receiver bandwidth a signal needs

Why it’s hard
Too little bandwidth closes the eye on its own; too much costs money and adds noise.
What I did
Sized receivers from the baud rate: the Nyquist frequency is half the baud rate, so 224 Gb/s PAM4 at 112 GBd needs about 56 GHz, and a 60 GHz sampling scope is adequate where the same data rate in NRZ would not be.
Result
Instruments matched to the signal, not over-specified.
08Industry · hands-on

Testing receivers with built-in electronics

Why it’s hard
A photodiode with an integrated transimpedance amplifier needs its channels powered and switched before it can be measured.
What I did
With my team on site, rewired the source-measure units and probe card for TIA testing, with a microcontroller as the communication bridge for initialising the amplifier and switching its four channels.
Result
All four channels switched and measured correctly on the device.
09Industry · hands-on

A low reading: tester or device?

Why it’s hard
A TIA channel showed little or no current; the cause could be the probe, the contact or the device.
What I did
With my team, proved the probe pins with aluminium and gold block tests, traced the low reading to surface oxide on the pads, and fixed it by repeated probing at the same spot.
Result
The channel read within the expected range; the tester was cleared, not blamed.

Part 3 of 4

The evidence

The on-wafer bandwidth chain I integrated, and the 224 Gb/s instrument plan I wrote.

3.1 · Setup and instrument plan

What the measurement looks like, and what it would take at 224 Gb/s.

On-wafer RF bandwidth test of a photodiode A vector network analyser drives an electrical-to-optical converter. Its modulated light travels by fibre to a fibre array at the edge coupler of a photodiode on the wafer. An RF probe picks up the photodiode output and returns it to the analyser, which measures S21. A reference optical-to-electrical module is used once to characterise the converter. Vector network analyserS21 SWEEP E/O converterMODULATED LASER Reference O/ECALIBRATION ONLY PIC WAFER · PHOTODIODE UNDER TEST PD FIBRE RF PROBE LIGHT RF OUT
The on-wafer bandwidth measurement I integrated: the analyser sees the whole chain, and calibration removes everything except the photodiode.

224 Gb/s PAM4: the instrument plan I wrote

  • RF 3 dB bandwidth · wafer, dieLightwave component analyser, 70–110 GHz
  • Eye diagram · wafer to module120 GBd BERT pattern generator + 60 GHz sampling scope
  • BER · moduleBERT pattern generator + error detection
  • BER · productionMulti-channel production BERT with clock recovery
  • Research patterns256 GSa/s arbitrary waveform generator

From my October 2025 survey with Keysight and Anritsu. I specified these; the eye-diagram and BER instruments were recommendations, not systems I operated.

Part 4 of 4

Why it matters to you

Bandwidth, eye and BER apply well beyond photonics.

Why it matters to you

From photonic wafers to any high-speed interface.

  • S21 bandwidth with calibrated de-embeddingCharacterising channels, fixtures and sockets
  • Eye height, width and jitterMargining high-speed interfaces
  • BER with PRBS patternsLink validation at production volume
  • Tester-versus-device root causeSeparating fixture faults from real failures
  • Instrument selection by baud rateRight-sized, cost-aware test cells