Development of Very Broadband Circuits

[mawyatt]

The thread about Cascode Amps got our Curiosity going and just discovered (wife was going to toss them in trash) some notes that we can share, many can't be shared tho

Long ago starting round ~2004 we began looking into alternate means to create miniature Electronic Warfare Systems, these "Systems" would eventually evolve into single-chip solutions, EW SoC so to speak!!

Significant efforts were put forth to develop Active replacements for the RF/MW/MMW combiners, splitters, mixers and so on that would cover ~DC to over 20GHz, later onto ~100GHz.

We began looking into core ultra-broadband amplifiers with controlled input and output impedances that could be altered into Active Input Combiners and Active Output Splitters, and later Active Mixers (later the PolyPhase Mixers were employed for many uses). Obviously for this effort to yield we needed to have good DR and low noise to replace the traditional components.

Initial work began with IBM SiGe BiCMOS processes (with plan to later into THz type InP) which supported SiGe BJTs over 300GHz Ft with 130 & 90nm CMOS. The core amplifiers were of the Cherry-Hopper origin that we added variable gain control by means bias current control. The designs are all Differential In/Out and utilized Quad Device & Layout Techniques to minimize process and temperature gradients.

1st is some quickly developed amplifier notes prepared back in March 2005 for some discussions & students. Note this is for a single ended version split in half by symmetry which made the analysis easier.

2nd is the test amplifier core, note the bandwidth peaking network at the top.

3rd/4th is the Gain control concept which is linear in dBV type control.

5th is the Gain Control schematic, note I1 and I2 for the core Amp.

6th/7th is response from 1Hz to 100GHz with AGC varied from 0 to 100ua.

8th is the Differential Input Referred Noise, note the 1/f characteristic on the left.

We also have some info on the Active Splitters/Combiners & Mixer if folks are interested.

Anyway, these were some of the core amp developments from back in 2004~5. These were fabricated in various SiGe BiCMOS processes and recall the tests performance was quite good, altho don't have any test data. Some fun stuff from way back then, and helped get our small research company acquired in 2006

Hopefully some folks may find these amplifier design concepts useful, they should apply equally well to any BJT, including the 2N3904 but with a much lower bandwidth.

Will try and answer any questions, if we can remember

Best,

[David Hess]

I am surprised you did not borrow some of the all NPN baseband RF IC designs from Tektronix.  I guess with 100 GHz Ft, you did not need things like Ft doubling and bridged t-coils.  Tektronix used Gilbert cells, sometimes they referred to them as paraphase amplifiers, for gain control but I think later they had something even better.

[TimFox]

A good reference book (albeit overpriced by Springer Verlag), much based on Tektronix work, is "Wideband Amplifiers" (2006).
https://link.springer.com/book/10.1007/978-0-387-28341-8

[mawyatt]

I am surprised you did not borrow some of the all NPN baseband RF IC designs from Tektronix.  I guess with 100 GHz Ft, you did not need things like Ft doubling and bridged t-coils.  Tektronix used Gilbert cells, sometimes they referred to them as paraphase amplifiers, for gain control but I think later they had something even better.

We were aware of the Tek stuff (the Ft doubler and the Ross inspired T-Coils), but ended up developing some things that went beyond that because of DR requirements and had access to Ft > 300GHz. For example, we had an active mixer based upon a Cherry-Hooper cell with a Caprio Quad within the internal loop and with Gilbert style current commutators which approached +50dBm OIP at <10GHz. As good as this was the PolyPhase Mixer became the mixer of choice for most advanced narrowband tracking work, placing the mixer up front with sub 2dB NF bare arsed to the world saying "Jam me if you can", that's pretty hard to beat with any technology with all the key features let along a single chip solution, and why DARPA created a special workshop on such, named Mixer First back in ~2011!!

Best, 

[mawyatt]

A good reference book (albeit overpriced by Springer Verlag), much based on Tektronix work, is "Wideband Amplifiers" (2006).
https://link.springer.com/book/10.1007/978-0-387-28341-8

Yes we had that book, very good reference, especially the mathematical portions.

Best,

[TheUnnamedNewbie]

Interesting to look at some of this stuff. I've also been involved in some high-speed circuits, though never DC up to the 100GHz, always bandpass (though >70 GHz bandwidth) at millimeter-wave frequencies.
Also always using FETs (either CMOS or SOI). Some colleagues of mine have worked in InP, GaAs, GaN, but really it isn't all that amazing. InP has amazing Fmax and all that, but the AM-PM is horrendous. Great at making amplifiers for sine waves, not so great at piping >100 GBit/second through

[BigBoss]

Interesting approach.
I have involved into few Cascode pHEMT MMIC wideband amplifiers while working but this is definitely difference approach.
I hope you share your experiences about Wideband Mixers and I/Q Modulators with us.
 

[mawyatt]

Will try and dig up some stuff on the other circuits (Mixers, Combiner, Splitter), if we can find anything. Have already looked into the special Cherry-Hooper based Mixer with the Caprio Quad mentioned but can't find anything, recall the very high Input and Output Intercept Points tho.

One of the issues we encountered that prevented us from using the older Tektronix type circuits mentioned was we were restricted to the relative low voltage SiGe BiCMOS processes initially, and needed everything to integrated so some of he classic techniques couldn't be applied (remember our goal was to move towards EW Systems at the Chip-Level.

This also led us into the Log Domain Filtering where the input signal is "Compressed" by the Log function then filtered, and after all the filtering then the signal is "Expanded" with the Exponential function. Can try and dig up some things on this if interested, think we have a few notes from the past.

Another outcome was on the EW Emitter side where we needed to generate arbitrary waveforms at RF/MW/MMW frequencies at moderate power levels (higher than possible with SiGe BiCMOS) and we looked in a composite GaN and Si device called a GaNsistor and DD2A (Direct Digital To Antenna), can also try and find stuff on this if interested.

Best,

[TheUnnamedNewbie]

This also led us into the Log Domain Filtering where the input signal is "Compressed" by the Log function then filtered, and after all the filtering then the signal is "Expanded" with the Exponential function. Can try and dig up some things on this if interested, think we have a few notes from the past.

I know logamps are used a lot now to predistort AM signals that will be detected with a square-law circuit (like power detectors and photo-diodes), but what are the benifits of doing this kind of 'log domain' filter? Are you trying to make non-liniear (pre)equalization?

[mawyatt]

This also led us into the Log Domain Filtering where the input signal is "Compressed" by the Log function then filtered, and after all the filtering then the signal is "Expanded" with the Exponential function. Can try and dig up some things on this if interested, think we have a few notes from the past.

I know logamps are used a lot now to predistort AM signals that will be detected with a square-law circuit (like power detectors and photo-diodes), but what are the benifits of doing this kind of 'log domain' filter? Are you trying to make non-liniear (pre)equalization?

No equalization, just a integrated replacement for a high DR tunable RF filter. These initial Log-Domain concepts were thought to have high linearity and be highly tunable at low supply voltages (power), however the self-generated noise within the Log Domain proved to be the achilles heal. These are quite complex and involve Current Domain operation as well as the Log-Domain which must realize tunable inductive type positive reactances as well as the usual tunable capacitive reactances. Will try and dig up some notes (maybe difficult to read tho) on these that were going to get tossed if you want.

Best,

Edit: Found some notes see attachments.

[mawyatt]

Additional images.

Best,

[mawyatt]

A few more.

Best,

[mawyatt]

Found something on the Cherry-Hooper Active Splitter and Combiner. Don't know why these images got flipped??

Also found some notes on a Dirac Impulse Sampler if folks are interested, had completely forgot about most of this stuff but starting to remember as it was headed for the trash


Edit: Fixed Images

Best,

[mawyatt]

Here's the noise performance of the combiner, very respectable. Note the low 1/f corner due to the SiGe BJTs, these have exceptionally good noise performance and have been utilized to displace other semiconductor technologies where phase noise performance was critical, but that's a different topic. The short circuit and 50 ohm noise plots indicate a ~40pa/Rt(Hz) input referred current noise density.

Best,

[RoGeorge]

Why does the noise density suddenly increase at higher than 10GHz?

I was expecting to see only some 1/f kind of noise increase at lower frequencies, but not at frequencies higher than the operation bandwidth.  Are my expectations wrong, or the noise at >10GHz is because something specific only to SiGe, or specific to this particular schematic, or to the log/exp inserted in the signal path, or maybe caused by something else?

[mawyatt]

This is the Input Referred Noise Density which means all the individual noise sources (current and voltage) in the circuit (resistor, BJT, MOS, Diode) are computed at nominal bias, then referred to the input by each elements associated gain (Root Sum Squared type summation).

Since each element noise contribution is gain referred to the input as the gain drops at higher frequency the individual element's noise contribution increases and the overall Input Referred Noise increases.

Should add that the Short Circuit Input Referred Noise Density plot as shown in #13 above is with the "Source Impedance" shorted, the other is with this impedance at 50 ohms. The Short Circuit noise is lower because the Input Referred Current Noise produces no noise voltage due to the short circuit source impedance, when the source is 50 ohms this noise current now produces an additional noise voltage as shown in the 50 ohm plot (Root Sum Squared type summation).

Hope this helps,

Best,

Edit:

Consider a simple amplifier with a Gain of 10, and an input device with 10nv/Rt(Hz) and an output device with 100nv/Rt(Hz), everything else in the amp is noiseless. The Total Input Referred Noise with be due to the input device 10nv and the output device referred to the input, or 100nv/Gain, or also 10nv. Then Root Sum Squared (RSS) is Root[(10nv)^2 + (100nv/Gain)^2], or 14.1nv/Rt(Hz). Here you can see that as the Gain decreases the Total Input Referred Noise density increases.

Now look at the output noise density. This would be the Total Input Referred Noise times the Gain, or 10*[14.1nv/Rt(Hz)] which is 141nv/Rt(Hz). We can also calculate this directly at the output as the RSS of Input Device Noise times Gain and the Output Device Noise. This will be Root[(10nv*Gain)^2 + (100nV)^2], or 141nv/Rt(Hz), which is same as using the Total Input Referred noise as it should be.

[RoGeorge]

This is the Input Referred Noise Density which means all the individual noise sources (current and voltage) in the circuit (resistor, BJT, MOS, Diode) are computed at nominal bias, then referred to the input by each elements associated gain (Root Sum Squared type summation).
...
Hope this helps,
...
numeric example
...

That helped a lot, thank you!

Since the plotted noise is the equivalent input noise if it were to consider our circuit as an ideal noiseless circuit (aka Input-Referred Noise), and since we have the gain vs. frequency response of our circuit as in IMG_3240.JPG, then it all makes sense.

To double check, by overlapping the Gain vs. Frequency plot with the Input Referred Noise Density vs. Frequency plot, we can notice the curves shape mirror each other horizontally (at higher frequencies), which means the output noise at >10GHz stayed about the same, as expected, so no sudden increase. 

[mawyatt]

If you take the example shown above and add another noise source of 50nv/Rt(Hz) in-between with a 5X Gain referred to the input, then the Total Referred Input Noise becomes 17.3nv/Rt(Hz). Then compute to Total Output Noise and note this added noise source Gain to the Output is 2X (10/5) and you get the Total Output Noise as 173nv/Rt(Hz) and if you take the Total Input Referred Noise {17.3nv/Rt(Hz)} times the overall Gain of 10X, then you get the same required answer!

Best,