F
Member
Anyone measuring a 40 GHz channel through a test fixture runs into the same wall: the fixture's own loss and reflections are larger than the effect you are trying to characterise. De-embedding is not optional at these frequencies. This post covers what we settled on after going through SOLT, TRL and AFR, with the numbers from our own bench.
1. Why SOLT is the wrong tool above 20 GHz
SOLT (Short-Open-Load-Thru) calibration is straightforward and the standards are cheap, but:
- The open standard's fringing capacitance is frequency dependent and hard to model above 20 GHz.
- The load standard's return loss degrades; a 50 ohm load that is 49.5 ohm at DC may be 47 ohm at 40 GHz.
- Characterising the standards requires a calibration kit model that itself must be validated.
In practice we see 0.3 to 0.8 dB of systematic error at 40 GHz with SOLT on a WR-28 waveguide setup. That is the same order as the connector loss we are trying to remove.
2. TRL is better but constrains the fixture
TRL (Thru-Reflect-Line) pushes the reference plane to the middle of the thru line and needs no fully known standards. It is accurate to 50 GHz and beyond, and it is what we use for on-wafer and waveguide work.
The constraint: the line standard must be a specific electrical length (typically 20 to 160 degrees of phase). For a PCB fixture with a 2.4 mm connector at each end, building a set of line standards that spans 10 MHz to 40 GHz means several lines. It is doable but the fixture becomes a project of its own.
3. AFR / 2x-thru: what we actually use for PCB fixtures
For characterising a PCB launch plus trace, the 2x-thru method (also called AFR, Automatic Fixture Removal) is the pragmatic answer:
1. Build a 2x-thru: two fixtures back to back, no DUT in the middle.
2. Measure S-parameters of the 2x-thru, 2 ports.
3. Split the 2x-thru into two half-fixtures mathematically (assuming symmetry).
4. De-embed each half from the DUT measurement.
The assumption of symmetry is the whole game. If the two halves are not electrically identical, the split introduces error that grows with frequency. We control this by:
- Fabricating the 2x-thru and the DUT on the same panel, same stackup, same copper weight.
- Keeping the two halves within 1 mm of length matching.
- Verifying symmetry by measuring the 2x-thru from both directions and comparing.
In our data, a well-built 2x-thru gives residuals below 0.2 dB to 40 GHz. A poorly matched one gives 1 dB residuals at 30 GHz and looks like DUT resonances that do not exist.
4. The step everybody skips: reference plane verification
After de-embedding, you must prove the reference plane moved. Two methods:
- Measure a known DUT (a 0 ohm resistor, or a short trace of known length) and compare the extracted delay against the physical length. At 40 GHz a 1 mm error is 7 ps, which shows up clearly in group delay.
- Measure the de-embedded thru; it should be 0 dB insertion loss and 0 degrees phase, flat across frequency. Any ripple is residual error.
We keep a "golden thru" in the lab specifically for this check. It takes five minutes and catches most de-embedding mistakes before they become design decisions.
5. Numbers from our bench (2.4 mm connectors, 40 GHz)
Fixture insertion loss, raw 3.8 dB at 40 GHz
After AFR de-embedding 0.4 dB residual
Return loss, raw -8 dB
Return loss, de-embedded -18 dB
Group delay ripple +/- 3 ps
Those residuals are the error floor of any eye diagram or TDR measurement made through that fixture. If your specification is 0.5 dB, a 0.4 dB error floor means you cannot resolve it. That is the real reason high-speed digital teams end up buying better fixtures rather than better algorithms.
6. Practical recommendations
1. Use TRL for waveguide and on-wafer work above 30 GHz; use AFR/2x-thru for PCB fixtures.
2. Always fabricate the calibration structures on the same panel as the DUT.
3. Verify the reference plane with a golden thru after every de-embedding run.
4. Report the residual error floor alongside your measurement; a number without an error floor is not a measurement.
5. Keep the raw (un-de-embedded) data. When a design fails at 35 GHz you will want to know whether it was the DUT or the fixture.
-- F-tone Networks, High-Speed Interconnect Engineering
Technical resources: https://www.f-tone.com
1. Why SOLT is the wrong tool above 20 GHz
SOLT (Short-Open-Load-Thru) calibration is straightforward and the standards are cheap, but:
- The open standard's fringing capacitance is frequency dependent and hard to model above 20 GHz.
- The load standard's return loss degrades; a 50 ohm load that is 49.5 ohm at DC may be 47 ohm at 40 GHz.
- Characterising the standards requires a calibration kit model that itself must be validated.
In practice we see 0.3 to 0.8 dB of systematic error at 40 GHz with SOLT on a WR-28 waveguide setup. That is the same order as the connector loss we are trying to remove.
2. TRL is better but constrains the fixture
TRL (Thru-Reflect-Line) pushes the reference plane to the middle of the thru line and needs no fully known standards. It is accurate to 50 GHz and beyond, and it is what we use for on-wafer and waveguide work.
The constraint: the line standard must be a specific electrical length (typically 20 to 160 degrees of phase). For a PCB fixture with a 2.4 mm connector at each end, building a set of line standards that spans 10 MHz to 40 GHz means several lines. It is doable but the fixture becomes a project of its own.
3. AFR / 2x-thru: what we actually use for PCB fixtures
For characterising a PCB launch plus trace, the 2x-thru method (also called AFR, Automatic Fixture Removal) is the pragmatic answer:
1. Build a 2x-thru: two fixtures back to back, no DUT in the middle.
2. Measure S-parameters of the 2x-thru, 2 ports.
3. Split the 2x-thru into two half-fixtures mathematically (assuming symmetry).
4. De-embed each half from the DUT measurement.
The assumption of symmetry is the whole game. If the two halves are not electrically identical, the split introduces error that grows with frequency. We control this by:
- Fabricating the 2x-thru and the DUT on the same panel, same stackup, same copper weight.
- Keeping the two halves within 1 mm of length matching.
- Verifying symmetry by measuring the 2x-thru from both directions and comparing.
In our data, a well-built 2x-thru gives residuals below 0.2 dB to 40 GHz. A poorly matched one gives 1 dB residuals at 30 GHz and looks like DUT resonances that do not exist.
4. The step everybody skips: reference plane verification
After de-embedding, you must prove the reference plane moved. Two methods:
- Measure a known DUT (a 0 ohm resistor, or a short trace of known length) and compare the extracted delay against the physical length. At 40 GHz a 1 mm error is 7 ps, which shows up clearly in group delay.
- Measure the de-embedded thru; it should be 0 dB insertion loss and 0 degrees phase, flat across frequency. Any ripple is residual error.
We keep a "golden thru" in the lab specifically for this check. It takes five minutes and catches most de-embedding mistakes before they become design decisions.
5. Numbers from our bench (2.4 mm connectors, 40 GHz)
Fixture insertion loss, raw 3.8 dB at 40 GHz
After AFR de-embedding 0.4 dB residual
Return loss, raw -8 dB
Return loss, de-embedded -18 dB
Group delay ripple +/- 3 ps
Those residuals are the error floor of any eye diagram or TDR measurement made through that fixture. If your specification is 0.5 dB, a 0.4 dB error floor means you cannot resolve it. That is the real reason high-speed digital teams end up buying better fixtures rather than better algorithms.
6. Practical recommendations
1. Use TRL for waveguide and on-wafer work above 30 GHz; use AFR/2x-thru for PCB fixtures.
2. Always fabricate the calibration structures on the same panel as the DUT.
3. Verify the reference plane with a golden thru after every de-embedding run.
4. Report the residual error floor alongside your measurement; a number without an error floor is not a measurement.
5. Keep the raw (un-de-embedded) data. When a design fails at 35 GHz you will want to know whether it was the DUT or the fixture.
-- F-tone Networks, High-Speed Interconnect Engineering
Technical resources: https://www.f-tone.com
