You have no idea how easy you have it

bag

June 2024


Alford manual tuner, circa 1980s

Focus automated tuner system for RF probing, circa now

There is no way to discuss this without sounding like some crazy uncle, but I don't care.  Over the past 44 years I have observed a lot of changes in the microwave engineering industry that have led to incredible improvements in engineering output.  No, we did not have to walk to work in a snowstorm, but we did wear neckties and pocket protectors and the kind of shirt that involves dropping them off at laundry.  I am allergic to starch so I was always a little more wrinkly than other engineers.  We did have a certain amount of "pin-up" girls in the lab, mostly in the lid of tool boxes so you could hide them when needed.

Much of the efficiency improvements follows along with the evolution in computers.  Electronic design analysis products have come a long way.   Who would have thought that a software company would be bought for $35B?  That just happened when ANSYS was sold to Synopsis. In 1980 the best you could hope for was time-shared copy of SuperCOMPACT, arguably the beginning of computer-assisted microwave design. You had to enter circuit elements by typing in the nodes they attached to, rather than wiring up a circuit using schematic capture.  It was easy to make mistakes by re-using a node number.   EM analysis software did not even exist.  Computers were time-shared using modems that passed data from telephone handsets at just a few bits per second.

I am a hardware guy so software topics bore me.  Let's move on, I want to focus on hardware improvements, in particular, laboratory test equipment.

Network analyzers

One of the biggest change in the past few years is that network analyzers have added a ton of functionality. This should not be taken as an advertisement for Keysight, but their PNA-X, configured with two sources and four ports, really does replace a rack of test equipment.  You can perform multiple calibrations and measurements without removing/replacing any connectors on your DUT.   Although I am not familiar with Rohde and Schwarz equipment, a quick perusal of their network analyzer brochure indicates they have many, if not all of the same capabilities.

For the PNA-X, hardware options include four RF ports, pulse modulators, a low noise receiver, direct digital syntheses and two sources that can be power combined.  A USB power head is used to calibrate DUT input power levels.  Noise figure can be evaluated using the cold source method (no noise source needed). Speaking of power heads (and noise sources), back in the day we had to manually enter the ENR for the noise source and the power factor for power heads, and interpolate it if we used frequencies that were not calibrated.  Now the equipment networks over the USB connections and simply knows these calibration parameters. In the old days, network analyzers didn't even have bias tees built into them, you had to to configure them externally.

Even if you buy the hardware for making all kinds of exotic RF measurements, you will need to purchase software upgrades.  Software options include time domain reflectometry, noise parameter measurements, pulsed RF parameters, third-order intercept, input/output power curves, phase noise, frequency converters, with or without internal LO signal.   I am not really doing justice to this remarkable piece of equipment, consult their website for more info. 

"Now that you have modern equipment", management asks, "can you clean out your lab of all the obsolete equipment?" Your answer should be, "somewhat".  In the past, noise figure measurement required a noise figure meter, a noise source and perhaps a converter.  Those items are now surplus!  You probably won't need to keep a family of power combiners that you used for two-tone measurements, but they don't take up a lot of room so don't chuck them yet.  Even though your network analyzer has two signal sources built in, if you have any separate sources, you should store them for times when your network analyzer is unavailable or when you need crude good/bad measurements in production.

Other things you need to keep:  any gear associated with high-power measurements, especially high-power laboratory amplifiers, high-power attenuators, couplers and power meters, and waveguide stuff.  It is possible you can jumper a power amplifier into the PNA-X, but if you don't know what you are doing you will blow up $500,000 worth of gear. I have never tried to do this, as the Hippocratic Oath states, the first rule is "do no harm".

Although the PNA-X does a lot of things well, it does not match the specifications for a good signal analyzer.  Once known as a spectrum analyzer, this gear has been taught many new tricks, such as noise figure measurement, phase noise, and all manner of modulation analyses.  From what I have seen, a high-end signal analyzer has a lower noise floor than its equivalent function in the PNA-X.

Network analyzer calibration

In the 1980s, coaxial calibration kits were available (7mm, type N). Mostly, "cable calibrations" were done to a 7mm sexless interface, using SOLT cal kit that included an expensive but fragile sliding load. Calibration would take maybe an hour, if none of the standards acted up.  7mm connections are very fragile and it did not take much abuse to cause measuerment woes.

If you wanted to move the reference plane up to your device, calibrating a network analyzer was something your company figured out for themselves.  Devices were always measured in fixtures, RF probing on wafers was not a thing.  You made some "known" standards, then developed some math to map the universe of impedances accurately across the Smith chart.  Where I worked, we made calibration standards based on a 400 mil long, 10 mil alumina through line and a set of off-set short circuits.  The reference plane we wanted to establish was at the end of a 10-mil alumina microstrip circuit with 200 mil length.  Then a device under test would be evaluated in a break-apart fixture that used the same style (but different) connectors and one half of the through line length (200 mil) on each side.  The device was mounted on a "center-bar" which was machined from brass or copper to be slightly wider than the device  you wanted to test, and of course gold-plated. It seemed like the machine shop ignored our instructons for "do not break edges" of the center bar, which resulted in a lot of scrap.  Try mounting a 20 mil wide device onto a surface that is less than 10 mils, then try wirebonding it...

We chose to use some cheap SMA flange-mount connectors, as the only precision connectors (7mm) were prohibitively expensive. The connector tabs were maybe 5 x 20 mils, intended for soldering down to wide PCB lines.  We cut them so they were tapered, then mounted them at 90 degrees to the microstrip lines which were ~9.5 mils wide.  We never soldered them down, this was all done with pressure contact.  We were mostly interested in X-band but at some point took data to 18 GHz. I saved the cal kit from the trash bin, it doesn't take up much room in my personal museum.  If I can dig up one of the DUT test fixures that went with it I will follow up with photo.  We used this mess to characterize GaAs MESFETs which had maybe 1 micron gate length and could muster 6 dB available gain at 10 GHz on a good day. The S-parameters that we gathered were used to successfully (to some extent) design MMIC amplifiers. Calibrating would take an hour or so if the frequency list was minimized (points spaced at 500 MHz or wider).  At the time, you needed two network analyzer/bias tee  set-ups, one for 0.1 to 2 GHz and one for 2-18 GHz. Learn more about old HP test equipment at the HP Memory Project.

Cioaxial calibration kit circa 1982, using off-set shorts and a through line with SMA connectors

By the 1990s, precision connectors such as 2.92 and 2.4mm were available, and so were precision mechanical-standard calibration kits.  Calibrations became more accurate, but it was still painful making and breaking all of the connections.  Today, cable calibration is routinely done with electronic calibration (E-cal) modules, thanks to Vahe Adamian.  All you need is a torque wrench and push a few buttons on your VNA.  Every once in a while I open an old 2.92mm coaxial calibration kit to see what is left of it, and see if it is suffering from putrid foam disease. Many of calibration standards are now scattered around the lab, used as adapters.  In the old days whoever was the perpetrator of such a crime would get taken to the woodshed.  Today, no one cares.  However, you should keep the short, open and through standard in good shape for checking if your e-cal is any good. 

Keysight four-port electronic calibration module

For device characerization outside of a coaxial interface, test fixtures were upgraded to spark-plug-style connectors by the 1990s and accuracy greatly improved as the connectors were more repeatable. But the real breakthrough in characterization was the advent of on-wafer RF probing, a topic for another day.

Data collection and processing

Back in the day, if you wanted to record spectrum analyzer or curve tracer data the only option was to use a Polaroid camera. Inventor Edwin Land was still running the Polaroid company up until 1981, although their peak employment occured in 1978.  He did not live to see his company declared bankrupt in 2001, passing away in 1991.  Polaroid is a case study in milking a technology all the way to the end. If you wanted to play tricks on co-workers you could take a picture of anything you want and leave teh film in teh camera so they would get an eerie double exposure.  I know what you are thinking, you have a dirty mond and we never did that.  Film was literally purchased and consumed by the case.

Polaroid scope camera, locked and loaded

For network analyzer data you could use a pen plotter.  The one shown below is fancy it includes a carosel for multiple pen colors.  A critical step in MMIC design, after measuring device S-parameters, is to fit equivalent circuit models to the data, which is used as the basis for amplifier predictions.  In the early 1980s this modelling was done by spending hours with SuperCOMPACT, optimizing the equivalent circuit to the measurement.  The equivalent circuit smoothed out the data, and permits removal the wirebond inductance and bond pads from fixtured devices. The accuracy of all of this was pretty sketchy, based on the crappy test fixures we used. Fit results were often plotted using a pen plotter.

Hewlett Packard pen plotter

Load/source-pull tuners

At the top of this page you will see an old-school, double-slug tuner, and modern tuner system.  I spent many hours with slug tuners, assisting in data collection for low-noise amplifier design, trying to get semi-accurate noise parameters.  The 1980s noise figure test set is shown below.  It needed a converter to hit microwave frequcnies, the noise figure meter only went to 1600 MHz.

1980s noise figure test set

The concept of noise parameters (Fmin, GammaOpt and RN) was not new, but determining them was stuck in the stone age. Here's a 1982 test plan, from memory.  There was a lot of math involved, I think we used Fortran, but I just took the data and someone else figured out how to process it.

  1. Perform a an offset-short test fiture calibration on  network analyzer calibration with the cheesy offset short cal kit, be sure to include bias tees.
  2. Insert the DUT in its fixure and take an array of S-parameters over frequency at different bias points that you think will give good noise figure results.
  3. Calibrate the network analyzer to coaxial ports using mechanical standards, to a 7mm recision interface (the slug tuner we used had 7mm connectors.) Be sure to configure the ports with external bias tees.
  4. Use math to calculate an "error box" for the input side of the test fixture, including its 7mm-to-coax adapter, all teh way to the wirebond interface.
  5. Set up a noise figure analyzer stand such as HP8970A, noise figure meter, 8771C converter box, HP 8350B / 83594A Sweep Oscillator (2-26.5 GHz),  including bias tees. Connect to equipment on an HPIB bus so it can communicate. Calibrate it so that a through line has zero dB loss and zero dB nose figure.
  6. Cascade the tuner and the device on the noise figure test set and bias it up to a nominal bias point where you already recorded S-parameters.
  7. Play with the tuner slugs until you have the lowest possible noise figure at a fixed frequency. Congratulations you now have data on Fmin at a single frequeny, bias point and at room temperature!
  8. Turn off the device (but remember tee bias point) and carefully disconnect the tuner.  Be sure not to move the slugs!
  9. Measure two-port S-paramaters of the tuner on the network analzer and store the data. Be sure to include teh measured NF in the file name!
  10. Put the tuner back with the device on the noise stand and turn it back on to the same bias point and frequency.  Now mix up the tuner slugs, the play with them until you measure 1 to 2 dB higher noise figure than Fmin.
  11. Disconnect the tuner, measure its S-parameters. Be sure to include the measured noise figure in the file name!
  12. Repeat steps 10 and 11 until you have four or five data points.
  13. Now to post-process the data.  Use SuperCOMPACT to cascade the tuner S-parameters with the error box and determine the reflection coefficient that the device saw for each measurment. also, use it to correct Fmin for the loss of teh tuner and the error box.
  14. Use some more math to fit a three-dimensional parabola over the data points on teh Smith chart, and calculate the effors between teh points and the fit.  Hopefull they are just a few tenths of a dB
  15. The parabola will reveal Gamma Opt, Fmin and RN if you did it right.
  16. repeat all of this nonsense over frequency and bias points until you are sick of it, or quit work and start a construction company or sell used cars.

What's coming? 

It seems like artificial intelligence is the hottest topic in microwave.  You would do yourself a favor by learning how to apply it to microwave engineering.  I am not gonna be a big help here, but maybe someone out there would like to contribute content on the future of AI in microwave design.

One controversial use of AI is for keeping an eye on worker productivity. AI surveillance is already a touchy subject, in the not-to distant future an AI algorhythem may decide who keeps their job and who gets riffed.  Once AI takes most of the heavy lifyting out of engineering, employees are are aready treated more like "resources" rather than people.  If you want to a say in this, you might consider forming or joining a union; collective bargaining will give you more of a voice on how AI affects your job.  If you thought that social media would only do good things for mankind, don't make the same mistake about the end state of  AI.

Check out the Unknown Editor's amazing archives when you are looking for a way to screw off for a couple of hours or more!

Fan/hate mail can always be sent to UE@microwaves101.com