HPM7177 ADC from CERN

[Kleinstein]

Price wise there is not much difference between the LTC2500-32 and the LTC2378-20. With the filtering done in the ADC chip one can get away with a slower interface. This can simplify things quite a bit and lower EMI issues.
The INL specs are quite good compared to most of the SD ADCs.
The nice point with a fast SAR ADC is that one knows where the difficult points are and one can thus do an relatively easy test on each sample.

Another point with SD ADCs is that the noise level can be higher at some evil levels:  SD ADC tend to be good on average or most of the time, but can have trouble with idle tones at a few points.  Reducing those idle tones is where the art starts.

[Castorp]

Price wise there is not much difference between the LTC2500-32 and the LTC2378-20. With the filtering done in the ADC chip one can get away with a slower interface. This can simplify things quite a bit and lower EMI issues.
The INL specs are quite good compared to most of the SD ADCs.
The nice point with a fast SAR ADC is that one knows where the difficult points are and one can thus do an relatively easy test on each sample.

Another point with SD ADCs is that the noise level can be higher at some evil levels:  SD ADC tend to be good on average or most of the time, but can have trouble with idle tones at a few points.  Reducing those idle tones is where the art starts.

Yes, it's all very true.

I'd say both types have advantages and disadvantages, but the lines are getting blurry. There was an old wisdom that 1-bit Sigma-Delta had to be inherently more linear, because the 1-bit DAC inside is linear by definition. Well, that's no longer true. These SARs have ppm and sub-ppm linear multi-bit DACs inside. Also, both types are realized using switched capacitors, so there are dynamic charge effects at the inputs and Vref pins (unless there are built-in buffers).

As for the idle tones - I really don't know how, but some designers have found a way to get rid of them in single-bit SD, or at least to suppress them way below the noise floor. I haven't noticed any of them in AD7177-2 (despite looking very carefully). In ADS1281 (4th order single-bit SD) they are present. I think they are even mentioned in the datasheet, in a rare display of honesty. In the old CERN Sigma-Delta they were known and famous. Every once in a while they re-appeared in different forms. Various types of dithering helped, but never fully removed them. It was an endless whack-a-mole game 

[Gerhard_dk4xp]

 The 2500-32 for example, unlike the 2378-20, doesn't support 1MSPS which can be very useful.

Where did you get that? The data sheet says differently and my lt2500-32 did not note it,
running all day long at 1 MSPS and 100 MHz SPI.

Data sink is a Beaglebone Black, collecting up to 3 such SPI ports in real time through
a PRU with some 32 bits -> 4 bytes conversion in the coolrunner2.

Analog source is a chopper with pV/rtHz noise and GaN switches.


Are there any 1/f data for LT2500-32?

Cheers, Gerhard

[MegaVolt]

Analog source is a chopper with pV/rtHz noise and GaN switches.

Could you say a little more about this source? What elements are used in it?

[splin]

 The 2500-32 for example, unlike the 2378-20, doesn't support 1MSPS which can be very useful.

Where did you get that? The data sheet says differently and my lt2500-32 did not note it,
running all day long at 1 MSPS and 100 MHz SPI.

My mistake, the DS says minimum DF = 4 but I'd forgotten about the no-latency mode that outputs raw 20 bit ADC data - it's been several years since I last looked at it in any detail.

Data sink is a Beaglebone Black, collecting up to 3 such SPI ports in real time through
a PRU with some 32 bits -> 4 bytes conversion in the coolrunner2.

Analog source is a chopper with pV/rtHz noise and GaN switches.

Cool. So why did you choose the LTC2500-32 over the LTC2378-20? It's quite a bit more expensive if you aren't going to use the digital filtering and has no other advantages that I can see, with the potential for additional internal noise if the DF can't be turned off completely (but you'd hope that it can be).

Are there any 1/f data for LT2500-32?

The datasheet almost certainly shows it in Table 2, best seen in the SSINC entries: noise drops linearly with SQRT(downsampling factor) until the filter bandwidth reaches 60.28Hz where increasing 1/f noise diminishes the noise reductions from further bandwidth reductions. Of course that could be down to factors other than 1/f noise but that seems to be much the most likely explanation.

But since you have the setup anyway,  haven't you measured the noise spectrum yourself with shorted inputs?

[Castorp]

The datasheet almost certainly shows it in Table 2, best seen in the SSINC entries: noise drops linearly with SQRT(downsampling factor) until the filter bandwidth reaches 60.28Hz where increasing 1/f noise diminishes the noise reductions from further bandwidth reductions. Of course that could be down to factors other than 1/f noise but that seems to be much the most likely explanation.

But since you have the setup anyway,  haven't you measured the noise spectrum yourself with shorted inputs?

I'd be interested in those results too.

The 1/f corner must be somewhere in the low Hz. If you look how noise goes down from 30.86 Hz to 7.53 Hz bandwidth ( factor of 4) - it's almost factor of 2. It goes from 270 nV RMS to 150 nV RMS.

Better yet - let's take factor of 16 difference. That's 460 nV RMS at 120.55 Hz bandwidth. Now, 460/4 = 115. We have 150 instead. It means if it's 1/f, it must contribute 96 nV RMS in the 7.5 Hz bandwidth (sqrt(96^2 + 115^2) = 150).

Doesn't look stellar if those numbers are right.

[Gerhard_dk4xp]

>> Analog source is a chopper with pV/rtHz noise and GaN switches.
> Could you say a little more about this source? What elements are used in it?

Something like this:
https://www.digikey.de/product-detail/de/epc/EPC2038/917-1138-1-ND/5774048

The +point is the small capacitance, leading to less charge injection than those
analog switches.
The minus point: needs at least 3V Vgs to really switch on and just 6 V Vgs to really die.
And soldering them onto the board was an adventure at 1 mm total size.

> Cool. So why did you choose the LTC2500-32 over the LTC2378-20? It's quite a bit more expensive
> if you aren't going to use the digital filtering and has no other advantages that I can see, with
> the potential for additional internal noise if the DF can't be turned off completely (but you'd hope that it can be).

you can get 24 Bit magnitude + 7 bit offset for every 1 MSPS clock on the SPI
and on the second port the filtered and decimated data in 32 bit format when they are "ripe".

At 1 MSPS, the first 660 nsec are for aquisition and communication should be avoided.
The other 340 nsec are just enough to fetch the 32 bits at 100 MHz.
I simply shoot them into a 32 bit shift register in the coolrunner and read that as 4 bytes into the BBB.


> But since you have the setup anyway,  haven't you measured the noise spectrum yourself with shorted inputs?

I'm not yet there. This is a voyage through pV/rtHz analog electronics, VHDL in the Coolrunner,
low level C programming in the BBB PRU processor, FFTW on the ARM and the GPIB/488-like server in Debian Linux,
socket communication on the LAN,  the app on the laptop under Linux Mint and Gnuplot for the result pics.

I can control the ADC over the chain   <  laptop-app / LAN / ARM-server / PRU / ADC  >  and get time series
of a few ksamples in the reverse direction, but the ping-pong buffer for large transfers works only for ping,
not for pong :-).  And sometimes I also have to do paid work.

And second opinions are always interesting.

Cheers, Gerhard

[Gerhard_dk4xp]

>  The 1/f corner must be somewhere in the low Hz.

For CMOS, that would be a reason for jubilation!


(I'm used to MUCH worse from my Agilent 89441A)

[Castorp]

>> Analog source is a chopper with pV/rtHz noise and GaN switches.
> Could you say a little more about this source? What elements are used in it?

Something like this:
https://www.digikey.de/product-detail/de/epc/EPC2038/917-1138-1-ND/5774048

The +point is the small capacitance, leading to less charge injection than those
analog switches.
The minus point: needs at least 3V Vgs to really switch on and just 6 V Vgs to really die.
And soldering them onto the board was an adventure at 1 mm total size.

Ah, ok, those pV/sqrtHz are input-referred. I wonder (for academic reasons) where's the 1/f corner, and whether the part works when cryocooled.

CMOS and low 1/f are not mutually exclusive, if it's the design objective. Like in autozero op amps, or some of these new high-resolution ADCs. It's no longer about having super-duper input stage transistors, it's more about how you do the chopping, autozeroing, correlated double sampling, or whatever trick it is.

[MegaVolt]

>> Analog source is a chopper with pV/rtHz noise and GaN switches.
> Could you say a little more about this source? What elements are used in it?

Something like this:
https://www.digikey.de/product-detail/de/epc/EPC2038/917-1138-1-ND/5774048

The +point is the small capacitance, leading to less charge injection than those
analog switches.
The minus point: needs at least 3V Vgs to really switch on and just 6 V Vgs to really die.
And soldering them onto the board was an adventure at 1 mm total size.

And what was the source of the signal with an ultra-low noise level?

[Castorp]

Gerhard - sorry, I didn't clarify that I was talking about the part you use for amplification, not the switches.

Yesterday I gave a talk on the HPM7177. It's a bit more pedagogical and oriented towards CERN-specific matters, but it does contain a bit of new characterization data from recent measurements, so it may be interesting for some:
https://edms.cern.ch/ui/file/2323037/1/tcc_1302_beev.pdf

[Kleinstein]

It is an interesting series of slides. It is a good point that the HPM7177 can be reference limited when measuring higher voltages.

The simple picture is to separate the noise in an additive part from the ADC itself and a multiplicative part, from the references. However it is not for sure that the ADC itself may not also have multiplicative noise, or higher noise at higher voltage levels. The simplest example would be noise of the reference scaling and buffer, but there could also be additional parts inside the chip.

I kind of miss a test if the ADC reading it's own reference (the 7 V from the LTZ) over a longer time.  I would expect a little more than the noise near zero.
The increase in the 1/f noise level from measuring a short to the 10 V sources (by more than a factor of 1.5) also suggests that the extra noise may be more than from the reference alone. With equal noise (both are LTZ1000 based) from the external and internal reference one would expect a factor of 1.4 .

[Castorp]

It is an interesting series of slides. It is a good point that the HPM7177 can be reference limited when measuring higher voltages.

The simple picture is to separate the noise in an additive part from the ADC itself and a multiplicative part, from the references. However it is not for sure that the ADC itself may not also have multiplicative noise, or higher noise at higher voltage levels. The simplest example would be noise of the reference scaling and buffer, but there could also be additional parts inside the chip.

I kind of miss a test if the ADC reading it's own reference (the 7 V from the LTZ) over a longer time.  I would expect a little more than the noise near zero.
The increase in the 1/f noise level from measuring a short to the 10 V sources (by more than a factor of 1.5) also suggests that the extra noise may be more than from the reference alone. With equal noise (both are LTZ1000 based) from the external and internal reference one would expect a factor of 1.4 .

You are absolutely right. There is a bit of "gain noise" - either from the ADC itself, or from the Vref scaling. I have measurements of own Vref (5 V, 7.1 V and 10 V - which is just the 5V amplified by 2x). I did not include them in the characterization report or in the talk, because the information is way too specific for people who are not deeply into the inner workings of these devices. Once I have a bit more free time (not sure when), I'll process all this data and I'll put it up somewhere - probably in the OHWR site.

In any case, this extra noise contribution is small, compared to the LTZ1000 1/f noise measured at 10 V.

I also have estimates of the LTZ1000-based PBC (John Pickering's 10mA/10V unit that we use here) and the Fluke 732B, obtained by calculating the cross-PSD of two digitizers that sample simultaneously. Once again - it's part of the tons of information that I presently keep for my own use. Not because it's secret, but because it takes time and effort to present and explain it.

[Gerhard_dk4xp]

> And what was the source of the signal with an ultra-low noise level?

Could be anything where you need  low noise with low source impedance:
power supplies, volt. references, control loops.


@ Castorp

That could be sth. like this:
<   http://www.hoffmann-hochfrequenz.de/downloads/lono.pdf     >

The design in that paper was handicapped by its too small
input coupling capacitor. That has been remedied with a
wet slug tantal. Now I get a clean 1/f response again also
with AC coupling. The noise of the op amp bias must be
shorted through the low impedance DUT, so the capacitor
must be MUCH larger than needed for f-3dB.

20 op amp inputs are a lot of shoulders to carry the load
of some abuse, but 4700uF++ @12V stores a lot of energy,
which requires some protection circuitry.

Sequencing takes some care, so there is an optically
isolated SPI interface now that also adjusts input, gain
and bandwidth via analogue switches.

I noted a rise of the equiv. input noise voltage above 500 KHz
which turned out to be skin effect in the U-shaped routing.
Updating that with a wire mesh removed that.

In a chopper that 1/f problem and the big capacitors do not
play a role. I have also tested a non-differential version of
the ribbon preamp in Art Of Electronics V3. That works as
promised and delivers 70 pV/rt Hz. There is a small smd version
also. The LTspice simulation is somewhat optimistic.

When you do cross correlation measurements, the input
noise current  produces a drop over the DUT resistance that
is common to both inputs and that does not average away.

Therefore I now prefer FETs. But the continuous reversal doubles
the input C for the apparent doubling of the input voltage
and the switching produces a large inrush current.
In a chopper, BJTs are probably better.

regards, Gerhard

[MegaVolt]

Could be anything where you need  low noise with low source impedance:
power supplies, volt. references, control loops.

I find it difficult to recall at least one source of constant voltage with noise less than nV. Did you use a low impedance divider?

[Castorp]

Thanks, Gerhard! That's some good work again.

I find it difficult to recall at least one source of constant voltage with noise less than nV.

These are usually very cold - Josephson junctions, SQUIDs, superconducting detectors. At room temperature - possibly some batteries. Some time ago I spotted some papers in which they used AD7177-2 for LF noise measurement of Lithium batteries. I can dig up the reference if someone's interested.

Another thing could be some narrowband matched circuit that transforms a higher impedance into a much lower one. There are such resonant detectors for image charges in Penning trap experiments. The people who work on them can usually afford good cryo-MOSFETs. Of course, they also don't care about DC and LF.

It's all pretty exotic in the end. After all - 1 Ohm at room temperature gives you 130 pV/sqrtHz. In liquid nitrogen (4 times colder) it's half of that.

[maat]

These are usually very cold - Josephson junctions, SQUIDs, superconducting detectors.

Unfortunately superconducting stuff has a rather large output impedance. We usually use a MESFET buffer, but at higher frequencies though, because their LF noise is horrible.

[MegaVolt]

These are usually very cold - Josephson junctions, SQUIDs, superconducting detectors. At room temperature - possibly some batteries. Some time ago I spotted some papers in which they used AD7177-2 for LF noise measurement of Lithium batteries. I can dig up the reference if someone's interested.

Thanks for the clarifications!!!

Link is very interesting. If not difficult, share it.

[Castorp]

These are usually very cold - Josephson junctions, SQUIDs, superconducting detectors.

Unfortunately superconducting stuff has a rather large output impedance. We usually use a MESFET buffer, but at higher frequencies though, because their LF noise is horrible.

What kind of device are you talking about?

It's true that "superconducting" doesn't mean it's a zero-ohm source. Some of them are close though. Single SQUIDs for instance - the only amplifier that's noise-matched to them is another SQUID. To clarify that - the input of a SQUID amplifier is a superconducting coil. I've worked on such a setup, with one SQUID at <100 mK and a large SQUID array at 4K. Eventually they demonstrated nearly quantum-limited energy resolution, i.e. a few times Planck's constant.

Another example is transition-edge sensors (TES). They can be either calorimeters or bolometers - depending on whether they measure the energy of single photons, or average radiation power. They are very low-resistance and very low-noise. It's basically a small resistance biased on the superconducting transition. It's a steep transition from 0 to a few Ohms, so they have lots of gain on that setpoint. Once again - you need an amplifier suited for very low impedances, e.g. a SQUID.

Sorry about drifting the topic far from ADCs. It just brings back some fond memories 

Here's one of those battery LF noise papers (seems the same author has many):
https://www.researchgate.net/publication/338209069_Electrochemical_noise_measurement_methodologies_of_chemical_power_sources

[Gerhard_dk4xp]

I did also some measurements on batteries.
<   http://www.hoffmann-hochfrequenz.de/downloads/NoiseMeasurementsOnChemicalBatteries.pdf     >

inspired from this NIST report:

<  https://www.researchgate.net/publication/3622215_Measurement_of_voltage_noise_in_chemical_batteries/link/555b4cc508ae6aea0816a420/download   >

It used to be on the NIST server of the timefreq group, but it seems it is moved around on a regular base.
The NIST people don't dare to disclose / endorse commercial products, but I wanted data on things I can
actually buy.  "I found a NiCd cell in my drawer" is not precise enough.

My preamp was better, but they had cross correlation, so they are better by a few dB over all.

The pretty amplifier from some posts above released its black magical smoke btw.
when measuring 4 Lithium cells. 

You see that wire from the op amp inputs over the DK4XP text to the "wrong" side
of the wet slug tantalum location. That was OK for verifying the skin effect, but was
forgotten to be removed before flight.    

"Opfer müssen halt gebracht werden"  -- Otto Lilienthal, early aviation pioneer
("Sometimes sacrifices are necessary")

@Megavolt: A LT3042 features 2 nV/rtHz noise density. I you want to verify that,
it's better to have >10 dB margin in your setup.

[splin]

I did also some measurements on batteries.
<   http://www.hoffmann-hochfrequenz.de/downloads/NoiseMeasurementsOnChemicalBatteries.pdf     >

It would be good if you could either rerun the tests or at least add a disclaimer at the start of the article that the results are all wrong below 100Hz due to the 1/f noise of the amplifier used.

https://www.eevblog.com/forum/projects/low-frequency-very-low-level-dc-biased-noise-measurements/msg655965/#msg655965

The results are interesting data for > 100Hz; below 100Hz they are still very useful showing the relative noise performance of the various batteries. Subtracting the estimated 1/f noise of the amplifier used gives a reasonable estimate of the actual low frequency noise of the batteries so all useful data.

I saw that you spotted the 1/f noise problem of the amp long ago but it's a shame you didn't get round to correcting your web pages detailing the amplifier and its performance.

[maat]

Unfortunately superconducting stuff has a rather large output impedance. We usually use a MESFET buffer, but at higher frequencies though, because their LF noise is horrible.

What kind of device are you talking about?

It's true that "superconducting" doesn't mean it's a zero-ohm source. Some of them are close though. Single SQUIDs for instance - the only amplifier that's noise-matched to them is another SQUID. To clarify that - the input of a SQUID amplifier is a superconducting coil. I've worked on such a setup, with one SQUID at <100 mK and a large SQUID array at 4K. Eventually they demonstrated nearly quantum-limited energy resolution, i.e. a few times Planck's constant.

Another example is transition-edge sensors (TES). They can be either calorimeters or bolometers - depending on whether they measure the energy of single photons, or average radiation power. They are very low-resistance and very low-noise. It's basically a small resistance biased on the superconducting transition. It's a steep transition from 0 to a few Ohms, so they have lots of gain on that setpoint. Once again - you need an amplifier suited for very low impedances, e.g. a SQUID.

Sorry about drifting the topic far from ADCs. It just brings back some fond memories 

Last bit of off-topic from me . You don't want to draw any current from the superconductor. You are only interested in a voltage fluctuation due an external exitation/disturbance. The superconductor itself has of course 0 resitance (exlcuding type II + magnetic field here), but any electron, that you remove from the sensor will disturb the measurement, hence add noise. Therefore your amplifier should ideally have infinite input impedance.

[Castorp]

I'm still wondering what kind of device you're talking about  Because the term "superconductor" is about as general as "semiconductor".

Sometimes you draw current, e.g. in a voltage-biased SQUID. Another example that should be familiar to the audience here is a microwave-irradiated Josephson junction, biased at a Shapiro step. It should be obvious why you don't want to buffer this voltage with any transistor.

[Castorp]

Kleinstein's post finally provoked my curiosity to look more carefully at the extra 1/f noise. Here's what I've got.

First - there's a multiplexer inside the HPM7177 that allows you to couple various internal voltages to the ADC. Those can be the +5V (which also goes to Vref), the bare unscaled LTZ1000 voltage (ca. 7.1 V), and a 10 V which is Vref*2. Here's how their noise spectra look like:



Now here's the integrated noise within one decade. I chose 1 to 10 mHz, as it seems to be fully in the 1/f region (also for 0V input). One plot line shows the noise on the internal voltages, the other one is with external 5V and 10V. Those come from 10 sources in parallel, feeding either into a 100 Ohm or 50 Ohm load. The external contribution is small (60 nVRMS/decade at 10 V, half of that at 5V), but I rms-subtracted it anyway. The result:



Now, one step further. I RMS-subtracted the 0V component. So what should be shown is just the voltage-dependent excess 1/f noise:



At the moment I don't have the time and patience to trace it all down, but a rough estimate tells me that the Noise Index of the Vishay bulk metal foil resistors should be somewhere between -40 and -50 dB, i.e. less than 10 nVRMS/dec/V, but more than 3 nVRMS/dec/V. The datasheet spec is for -40 dB. If they are actually much better, i.e. better than -50 dB, then this extra noise must come from the AD7177-2.

It's actually quite a jigsaw puzzle to separate the components. There's one part in the input attenuator and another in the Vref scaling. When you measure the +5V internal voltage, the component from Vref scaling won't show. That could possibly explain why the excess noise doesn't scale linearly with voltage.

Anyway - it looks like it's not as negligible as I thought before. I think I had a wrong number stuck in my head - perhaps I confused the 7.1 V measurement for 10 V. Or maybe I hadn't subtracted the contribution of the 10 V external source that I used before. In any case - the results are still pretty good.

Edit:

For completeness, here are the spectra with external 5V and 10V. Some plots look thinner than others because they were averaged over much longer time (e.g. one week instead of overnight). Also the EMI pickup varies due to the cabling, etc. It doesn't matter for this particular study of 1/f noise. Also, the 50 Hz spike may look big, but actually it isn't. It's the FFT bin width which is very small (0.001 Hz), which results in a large processing gain for narrowband signals. Anyway, those details are not relevant.



Once I tried to measure the excess noise in bulk metal foil resistors. Back then I concluded it must be better than -40 dB NI, which was my measurement limit, and it was sufficient at that time to point my attention elsewhere. It would be nice to solve this mystery eventually.

[Kleinstein]

AFAK some of the AZ OPs also have a little bit of 1/f noise. So it is not only the AD7177 and the resistors as possible sources. I would expect the foil resistors to be much better than -40dB noise index. I would be more surprised if the resistor contribution is significant at all.

Having more noise from the ADC and scaling means the noise contribution from the LTZ1000 reference is a little smaller. The overall noise has not changed, only a slight change in where the noise comes from. The reference still is the largest contribution.

For the really critical parts it would be more useful to have 2 completely separate units in parallel and not use one ADC with 2 x LTZ1000 for the reference. Looks like this is the plan anyway.