How are the ADF synths so good?

[ezalys]

Looking at the phase noise on the ADF4372, we see -117 dBc/Hz at 4 GHz 100 kHz offset. The V600ME14 features -115 dBc/Hz at 3 GHz 100 kHz offset. This VCO is common in numerous instruments. Looking at VCOs from synergy microwave, others from zcomm, you usually see the ADF beating others by a narrow or somewhat larger margin. What’s going on? Am I missing something? How is this integrated device so good? Has it largely obsoleted other VCOs in phase noise performance?

[KE5FX]

Yeah, you don't want to be in the discrete VCO business at this point.  A few years ago the RFIC designers started building what amounted to HP 8662A VCOs on a chip -- meaning they began using lots of switched inductors, rather than trying to tune a single inductor across a wide frequency range with a single varactor.  That seems to have been the key.  The reduction in tuning sensitivity was massively beneficial, noise-wise, although it complicated the self-calibration process.

There is no reason the discrete VCO people couldn't do the same thing, of course.  But it wouldn't be quite as simple because tank-circuit selection and calibration would have to be handled by the customer rather than on-chip logic. 

If I were running a VCO product line, I'd abandon that playing field in favor of tunable filters.  There are few reasons to use a discrete VCO anymore, but the few RF filter ICs that have been produced are abusively priced.

[ezalys]

Are there some parts besides the ADF that I should have a look at?

[Gerhard_dk4xp]

LMX2594/95.
Or the TI app note on "Wilkinson-combining" multiple synthesizers, but
that hurts in the pocket after the first few dB. ( LMX2820 )

<    https://www.ti.com/lit/an/snaa361a/snaa361a.pdf?ts=1702059134500     >

Gerhard

[KE5FX]

Are there some parts besides the ADF that I should have a look at?

'ADF' covers a lot of ground.  Only the least-noise sensitive and most cost-sensitive applications would want to use the original ADF4350, but the ADF4377 seems to be at the top of the leader board at the moment, along with the LMX2820. 

Nothing wrong with the LMX259X parts, though. 

[ezalys]

I guess I’m curious how Keysight is doing it these days. Are YTOs still in style?

[KE5FX]

One, maybe two more generations' worth of improvement at the current pace will be enough to shut YTOs out of the market in favor of on-chip microwave VCOs.  They aren't quite defeated yet but the writing's on the wall.

YIG oscillators were great for a specific niche: instrumentation that needed oscillators that could sweep across wide frequency ranges and still provide low broadband noise in narrow sweeps.  Those block diagrams have been replaced by SDR architectures for the most part.  The local oscillator in a modern spectrum analyzer may need to hop anywhere from a few dozen MHz to several hundred MHz at a time, but it won't be called upon to sweep continuously.  The analyzer will simply digitize the IF at each hop and stitch the plot together as needed.

High-end spectrum analyzers probably don't use off-the-shelf integrated VCO/PLL chips, because they still care a lot about broadband phase noise.  (To be clear, you can get superb narrowband as well as broadband PN performance from those chips, but it's not just a matter of plopping a single chip down on the board.)  They are most likely using custom PLL synthesizers built with discrete VCOs.  If they do still use YIGs they won't for much longer, and I wouldn't bet the farm on the discrete VCOs either. 

VNAs still need to sweep their stimulus signal, on the other hand... but they care much more about speed than broadband phase noise, so that's more bad news for YTO fans.

[ezalys]

How about a signal source like an MXG, supposing you just need a single frequency? I guess I don't see any off-the-shelf discrete VCOs with phase noise that beats the ADFs. Are these custom designs that aren't available off the shelf? Are there some good documents on how they're designed?

[KE5FX]

There are VCOs, and there are VCOs.  The ones that have been obsoleted by the integrated VCO/PLL chips tend to be parts with wide tuning ranges, but there are still places for narrowband VCOs in reference-locking loops and such.

The trend in signal generator design is to do as much as possible digitally, at baseband, then translate the spectrum up to the desired carrier frequency.  Same deal as when designing a receiver or spectrum analyzer, but in reverse.  If you have hundreds of MHz worth of baseband bandwidth to work with, you don't need either fine frequency control or sweep capability in the carrier-generation stages.

Chances are that the customers for the high-dollar Keysight boxes want vector signal generation capabilities anyway, so the older block diagrams for generating a single CW or modulated signal at a specific frequency are less useful now.  So the designer might start by using a narrowband SAW VCO to establish a clean reference signal near 500 MHz, then feed that to a comb generator, then lock a wideband VCO to one of the comb lines.  The VCO specs in that case would be less critical because the lock bandwidth can be very wide.   

Someone building a 10 kHz - 60 GHz generator, for instance, will probably do it by mixing the output of a couple of these loops, each of which will have multiple VCOs for different parts of the tuning range.  The synthesizer just has to provide a clean signal within a few hundred MHz of the desired carrier frequency.  A vector modulator will handle the rest of the job, usually at a stage prior to the final mix. 

As for the VCOs themselves, Keysight would be likely to use their own designs.  Mid-tier manufacturers might put some ZComm or Synergy parts on the board.  At the lower end, you'll see an LMX2595 (IIRC) in the Signal Hound VSG60A, which will likely outperform many of the legacy designs used by the aforementioned mid-tier OEMs. 

The lowest-cost vector signal generators might use the integrated SDR chips from Analog Devices or Lime Micro or somebody like that.  I can't think of anybody in that category, commercially, but I've used the bladeRF as a VSG plenty of times, especially before the VSG60A came out.

Definitely should check out the Signal Path channel on YouTube if you want to see what is being done in late-model signal generators.  Sometimes he'll post a block diagram and/or label the actual components on the PCB.

[radar_macgyver]

One can still outperform an integrated VCO using for example a coaxial resonator oscillator (CRO). These usually have tuning ranges of only a few MHz. I use them in clock trees, for example, or to generate a reference for an LMX2594 so that it can operate at or near its 300 MHz PFD max. Another topology is to use a PLL with a CRO to generate a clock for a DDS. If you think about it, the multiband integrated VCOs achieve their performance by having a narrow tuning range per band, and hence a low kVCO. A CRO can achieve a higher Q than an on-chip LC tank, and can achieve a lower phase noise for a given kVCO. Similarly, one still finds uses for dielectric resonator oscillators (DROs) for fixed frequency applications.

The ADF4377 is just astonishing though - 500 MHz PFD max and 10-15 dB better than an LMX2594.

[G0HZU]

Yes, it's possible to make really good narrowband oscillators using just a few cheap parts.

The link below shows a youtube video of a simple 1.5GHz free running oscillator I designed recently. This uses a transmission line resonator and the aim was to make something very stable and with low phase noise.



I've used very light coupling to minimise loading for the stability demo. It hardly drifts at all in a couple of minutes.

At about 2kHz offset, the phase noise is about -100dBc/Hz.
At work, I gained a lot of experience designing VCOs in the 1990s. I designed narrowband and wideband VCOs back then. However, the NRE costs are significant for this stuff (especially for low volume stuff) and so at some time around 1998-2000, the company decided to only use commercial VCOs.

The various ADF and LMX PLL/VCO chips really are amazing. I've recently used an LMX2572 in a design at work. This sort of performance was unthinkable back in the 1990s. Back then I was wrestling with the limitations of using PLL chips with an external dual modulus prescaler (and possibly another prescaler ahead of the DM prescaler). The aim was to make an octave wide PLL up in the GHz region. I can remember using A Zcomm C706 VCO in an early version. This tuned from about 600MHz to about 1200MHz. Later versions tuned up to beyond 3GHz and I managed to escape from having to use external DM prescalers as the PLL/prescaler chips became faster and were integrated into one device.

[KE5FX]

That looks great.  It used to take a non-trivial amount of hacking to get a 5 MHz LC oscillator to remain within 100 Hz for any length of time, but you're in that ballpark at 1500 MHz.  What's the tuning range like?

[coppercone2]

I wonder if there is some super YIG in the works some where. They must have figured something out after 40 years.

I heard a few stories related to electronics shmoozing around 'brains' about some very impressive variant prototypes of certain technologies deemed to be weak/older that had some crazy performance increases but were otherwise limited by some manufacturing issues, thermal design, etc. The obstacle was cheap mass production. R&D get cut early if they can't provide a reasonable promise of solutions within some projected budget...

the YIG is totally fascinating as an object and that it uses physics effects. i wonder what else you can get from it. The PLL and Quartz etc are very boring in comparison. meta material yig spheres?

[G0HZU]

That looks great.  It used to take a non-trivial amount of hacking to get a 5 MHz LC oscillator to remain within 100 Hz for any length of time, but you're in that ballpark at 1500 MHz.  What's the tuning range like?

Thanks. It doesn't really have a tuning range as it was just designed to sit at one frequency. However, it could be modified to become a varactor tuned VCO. I don't think this would add much noise as long as the tuning range was something like +/- 10%.

I've got a few fairly modern (high end) spectrum analysers here that still use a YIG for the LO. Also a couple of signal sources, including a low phase noise vector sig gen that has a YIG based LO. This was a very expensive sig gen when new almost 20 years ago!

These analysers are all <20 years old. The newest analyser I have here (with a YIG based LO) is about 11 years old.

At work there are still a few PXA signal/spectrum analysers in the labs and these use a YIG oscillator. However, the PXA has been around for a few years now.

[Tony_G]

Sorry for the OT post...

Jeremy, please check out - https://www.eevblog.com/forum/repair/g0hzu-e3348c-psu-repair-cant-pm-reply/msg5217057/#msg5217057

Sorry I typo'd the model number, you might have missed the post.

Thanks,

TonyG

[G0HZU]

Hi TonyG
Thanks for this. I think I've sorted out the inbox issue and sent you another PM.

In terms of synths and alternatives to the excellent ADF chips, I've borrowed the official Ti LMX2572EVM eval board I'm using at work. I'll try and get it working again as it was fudged into another test fixture for a while. Sadly, I don't think any of my test gear can measure how good it really is. I don't have any formal phase noise measuring gear here. I've only got some decent spectrum analysers and a crystal notch filter and so I can measure the LMX2572 at a few frequencies...
Regards
Jeremy

[Mr Simpleton]

The ADF synths were always my go-to when needing a signal source, but the old Hittite (HMC833 family) did deliver better phase noise. Then I disovered the LMX2594  As far as I can tell it still is on top regarding phase noise, and the (nice) problem one run into is to source a good-enough reference. One have to start looking at Wenzell offerings, and then there is the problem on validating the numbers. Can be pretty hard if you do not have access to a good PN-test set. And I especially like the 15GHz  LMX25954 ability to go down to 10 MHz!. Oddly the Analog ADF often stops at >50 MHz.

The early versions of the 2594EVM did include a good  TCXO100 MHz reference, but sadly this is missing on the one sold now. Instead they have an integrated clock source that does lack a bit. Using a high performance 10 MHz reference, lock one LMX2594 and have this output 100 (or 200) MHz as a reference to the main synth should give really good performance and and also by adjusting values, reduce spurs from the main loop.

The number of registers (112!) can be a bit overwhelming when trying the LMX2594, but by using the TI software you see exactly what is written, where... and by using this information one can easily see what registers that are affected when changing settings. 

[Gerhard_dk4xp]

The ADF synths were always my go-to when needing a signal source, but the old Hittite (HMC833 family) did deliver better phase noise. Then I disovered the LMX2594 
...
The early versions of the 2594EVM did include a good  TCXO100 MHz reference, but sadly this is missing on the one sold now. Instead they have an integrated clock source that does lack a bit. Using a high performance 10 MHz reference, lock one LMX2594 and have this output 100 (or 200) MHz as a reference to the main synth should give really good performance and and also by adjusting values, reduce spurs from the main loop.

The number of registers (112!) can be a bit overwhelming when trying the LMX2594, but by using the TI software you see exactly what is written, where... and by using this information one can easily see what registers that are affected when changing settings.

Has anybody written a software driver for the LMX2594 for generic SPI?

I have made a tiny board with an 2594 that can be controlled by TI Software
and the TI USB I/F board. When I had the board working, the chip was no longer
available, so there was no real reason to follow that further.  Now the availability
has improved.

I'm working on a wide band down converter to convert everything up to
12 or 20 GHz down to 5...28 MHz. There is an ADF4002 to lock an ECS-2522
100 MHz oven (Digikey) to an external10 MHz ref.
Abracon 100 MHz non-ovenized should also work.

The ADF4002 is in the middle/left, the oven right to it. Since it delivers
3V3 CMOS, I simply buffer it with 2* LVC125, filter out the 300 MHz harmonic,
SKY MMIC for same gain & squaring, two 900 MHz SAW filters à €1.50 to clean up,
1 more MMIC and the LO for the 2nd down conversion is done.

On the right side is the home for an LMX2594/95. It currently gets its clock from
the 100 MHz oven. 900 MHz from LO2 should be possible, too, but I did not want too
many unknowns right from the start.

The LMX is the 1st LO. There is a low and a high band, nervous DC.. 3GHz and
3 GHz to 12 or 20, depending on LMX and available mixer. I wanted a mixer
to 20 GHz with built in driver, and you guess, it is unavailable now. The alternative
is a Hittite HMC-220, which results in a hole from 3-5 GHz and ends at 12GHz.   :-(
The 1st IF is 900..928 MHz with the same SAW filters.

On the top right is a LT3045 regulator for the 3V3 with enough power for the oven.

The idea is to use 2 of these boards to extend my Timepod from 30 MHz to
wideband with cross correlation. Should also work for ham radio and deep space
network with the right preamps/filters.

I'm still soldering, the 900 MHz LO2 works already. :-)
The board is 4 layer FR4 from JLCPCB. for $5. 10*9 cm. I have 5 of them.
It will be open source. Comments / proposals welcome.

Gerhard

[KE5FX]

Has anybody written a software driver for the LMX2594 for generic SPI?

There are a surprising number of candidates on Github, actually -- I don't know if they'll come up if you search on the site directly, but if you Google lmx2594 site:github.com you'll see a bunch of options.

[G0HZU]

I've not used the 2594 but I have some experience with the LMX2572. I wrote the ICD for the one I used at work but I didn't do any of the actual programming code. There are about 126 registers in the 2572 but most of them can be configured with a reset. So only a handful of registers need to be programmed for general use if you just want to use it in integer N for example.

For dev work I used an FTDI 4232H module to interface a PC to the LMX2572 but there's lots of cheaper options like an Arduino or a raspberry pi if you already have one.

I've still got the official LMX2572EVM board here with the TICS PRO SW. It should be easy to sniff the SPI to see what it does and the TICS PRO SW can do a register dump as well. Something like an Analog Discovery /2 /3 would be an ideal tool to look at the SPI.
When I did this for the ICD I noted that the TICS PRO SW writes to all 126 registers when it sets up the default 3GHz output frequency.

However, I think it's possible to just do a reset and program a handful of registers. You also have to make sure you calibrate the VCOs via the register settings. It's an incredibly versatile chip, but it should be fairly easy to control it (in integer N mode) via an Arduino for example.

[radar_macgyver]

The recommended way (per TI forum posts by Dean Banerjee) is to use TICS pro to generate a set of registers and load these when your micro starts up. Then, change the few registers you need, for example, the N divider values, output dividers, etc. The power-on defaults may not be optimal, and there are a number of registers that are undocumented, but whose value affects the behavior of the chip. Using TICS pro will configure them appropriately. As @G0HZU mentioned, TICS can generate a register dump after you've configured things the way you like. I wrote a simple script to turn the register dump into C source that I could include into my driver for the initial setup.

Please see attached for my LMX2594 driver. My application has three LMX2594s, driven by an LMK04832 clock generator, so the reference frequency into the LMX2594s depends on how the LMK04832's output dividers are configured. This is stored in g_ref_freq[chn]. The code has a bunch of application-specific stuff, but the part that may be of interest are the output divider configuration and the VCO calibration assist - this helps to reduce the time taken to acquire lock after large jumps in frequency. I didn't use symbolic constants for register names, etc. but with datasheet in hand, you can figure out what the code's doing. The target is a PSoC5, running FreeRTOS.

FYI, the LMK04832 was used to lock a 100 MHz VCXO to a 10 MHz GPSDO, then locks a 3.2 GHz VCO to the 100 MHz VCXO and divides that down to provide reference clocks for the LMX2594s. The PFD frequency for the LMX2594s is 290.9 MHz (3.2 GHz/11), but can be changed if this results in an integer boundary spur near my target frequency. The overall results were too low to measure directly on an N9020A MXA with a phase noise measurement app. Indirect measurement by using the LMX2594 output in a radar system LO indicated excellent performance (~50 fs RMS jitter), better than what we could achieve with a COTS synthesizer costing $$$s.

[Mr Simpleton]

When I was testing the 2594EVM, my intention was to generate a finite number of frequencies. So I started with the TICS program, and generated the settings for the needed frequencies. Hooked up a simple logic analyzer and took note on how the data were sent over the spi bus. As it turned out only ~10 registers were affected once the initial set-up was done. I then transfered the data to an ESP32 board (Arduino compatible) board, and used it to load the registers, frequency was then controlled using a simple terminal program and USB.

With this simple set-up I could also have the ESP32 to output a strobe signal which was used to trig a Tek RTSA, and made it really easy to monitor how long the 2594 would take to change frequency, once the registers were uploaded. In my case I think it was ~180ns. Of course a large jump would mean a longer time. Still pretty impressive.

But due to other factors the project was cancelled, and my 2594EVM was put away.

There are some commercial products that uses the 2594, ValonRF is one which also have TCP/IP interface and have one 2594 to generate the reference for another 2594 (WindFreak and ERAsynth are also using 2594)

BTW I have been using a PXA N9030A with phase noise option, and the outcome is just a plot of the internal phase noise of the spec... the LMX2594 easily out-performs the N9030, the 2594 was locked against a Wenzell 100 MHz OCXO meeting the close-in phase noise mentioned in the data sheet.

[G0HZU]

I fired up the Analog Discovery 2 and had a go at sniffing the SPI bus feeding to the LMX2572 PLL chip. The register grab below was taken after sending the default settings to the LMX2572 via the TICS PRO SW.
Also included is the  register dump data using the TICS PRO SW.

I've put them both side by side in an excel file as attached. You can see that the TICS PRO dump and the AD2 SPI grab both agree for all 126 registers.
The first 8 bits are for the address and the remaining 16 bits are data.