Xitron 2000M (Teardown)

[pigrew]

The Xitron 2000 is a 6.5 digits calibrator, designed in the late 1980s. It's a very cute box which I really want to love, but I'm have having relationship issues with it.

The calibrator uses a LM399A voltage reference, driving a 3-bit PWMDAC (244 Hz at about 4 kHz), yielding a 24-bit resolution. The DAC voltage is routed through a few amplifiers range switches, and a transconductance amplifier to provide 22V, 2.2V, 220mV, 22mV, 22mA, 2.2mA, 220uA, and 22uA ranges). At its lowest ranges, it has resolutions of 10 nV and 200 pA. Calibration constants are stored for (-), 0, and (+) for each range. An "internal calibration" is also provided so the user can periodically null the output (using a built-in ADC). The lowest-bit size is also adjustable via an internal trimmer resistor. There is no adjustment of linearity.

My build has two outputs: fairly normal banana binding posts on the front and huge copper blocks on the back which are intended to be used with thermocouple wire. Beware that the manual insists that "Cd18" (Sn50Pb32Cd18) solder be used on the output terminals (so use a fume hood or go outdoors as needed).

My unit was very dead upon delivery. As it turns out, the SLA battery had achieved a negative voltage, and the power supply won't startup without a battery (or large capacitor bank) in place (1.5A @6V inrush current at startup). Figuring out that it needed a good battery for operation took me a while to figure out, during which case I applied power to the PSU without loads being in place, and blew up a pair of tantalum output capacitors (NB: Don't power the power supply without a load in place). The power supply uses a linear regulator for the digital circuits (5V) and a flyback for generating isolated rails for the a analog board. A series of optoisolators lets the two boards communicate.

The microcontroller is a Motorola MC68322 (firmware in EEPROMs), and has a TI TMS9914 GPIB chip for instrument control.

One striking thing about it is the lack of ceramic capacitors (there is one: The Y capacitor across the flyback transformer): Poly films are used for decoupling digital chips, mica are used on the voltage reference board, tantalums are used for bulk. This isn't a bad thing, but it seems overly expensive. Maybe they were worried about ceramic capacitors being piezoelectric and creating noise in the circuits?

Anyway.... the issues. There are quite a few things that would need to be corrected to get it up to good condition. I'm not sure if it's worth improving to keep, or maybe I'll just resell it on eBay.

Battery overdischarge
The battery circuit is designed to disconnect once a minimum voltage is reached (measured to be 4.46V on my unit). It doesn't work. The circuit's parasitic above the threshold is 4.2mA. Once it reaches UVLO, it drops to 2 mA until it goes to zero at 3.5 V. This is due to LM339 comparator IC2 not going high-impedance when its Vdd is below its operating range. I believe this is what caused my negatively charged battery. I fixed this by adding a BSS138 in series with the comparator output.

EMI from the flyback!!!
Putting an oscilloscope probe anywhere near the analog board shows pronounced 100 kHz switching noise with 27 MHz ringing. This also couples onto the output terminals. Adding a snubber onto the primary of the flyback likely would help. Another option would be to replace the flyback with COTS switching modules (although I was having trouble finding modules that can output +/-27.5V). The attached spectrum analyzer image is with the output directly connected to 50-ohm coax to a spectrum analyzer. It shows quite a few frequencies at powers of about -77 dBm. This is on the order of 30 uV_RMS. Spurs from the clock and switching regulator are visible.

Distant signal return paths
The digital signal ribbon cables (both digital board to PSU and PSU to analog board) have no ground wires. This creates a huge loop which transmits an always-running ~0.9MHz clock to the calibrator's output. Adding extra ground wires near the signal cable should help, but is invasive. Perhaps adding a small resistance in series with the outputs would be good enough.

Difficult mechanical design for repair
Dismantling the chassis is needed for battery replacement. It's very difficult to get to the tiny nuts holding the chassis together. The analog board must be desoldered from the output terminals for many repairs.

LM399?
It's not a LTZ1000A, but is likely good enough for 6.5 digits. I don't have any complaints, though some small tweaks are possible. The Zener current is 930 uA (by my calculation), but maybe it should be increased to the standard 1 mA. As another tweak with negligible effect, I'd power the heater with +/-15V, instead of just the positive rail, so that ground currents are reduced. Adding a series heater series resistor wouldn't be a bad idea, either.

Reference op-amp stability

Can a LM308A really drive a 10uF tantalum capacitor as a load? I wonder if it's close to oscillation. LTSpice simulation shows damped oscillations at 12 kHz.

Linearity and temperature sensitivity
The PWMDAC switches between the reference voltages using high-resistance CMOS switches (AD7512). These switches are about 60 ohms, but that varies with Vds and temperature. This switch is in series with precision wirewound 100 ohm resistors. These switches are definitely the limiting factor for temperature sensitivity and linearity. Examining the schematic, there is a feedback loop from DACOUT back to the switch node. At first, I hoped that this was for charge balancing (to reduce ripple), but I'm thinking it's to compensate for the non-linearity of the switches. These switches should be replaced.


The oodles of precision wire wound resistors, poly-film caps, cadmium solder, and tantalum caps suggested to be that this should be a rock solid, well designed machine. Unfortunately, the electrical design lets me down. I think it'd work well for calibrating 4.5 digit multimeters, but is questionable for its intended use of supplying sub-millivolt signals to thermocouple systems, or nanoamperes due to excess EMI.

(At this point in time, a few units on eBay have recently sold for US$110 to $160 each. New units are for sale with a MSRP of US$3,880.00. I'm assuming there were design updates, perhaps correcting the above-mentioned flaws.)

[pigrew]

Specifications

[dietert1]

Thanks for the review. Did you understand how they use the two helper switches in IC9? Could be a similar scheme as in those Datron PWMs where they add in +/- 2 mV from a helper PWM.
For ripple compensation it is easy to add, e.g. https://www.edn.com/cancel-pwm-dac-ripple-with-analog-subtraction/.
Those switches are special in that they have very good resistance match (1 % typical), which is about 0.6 to 0.7 Ohm. With the 40 K PWM resistor this amounts to an inaccuracy of about 4 or 5 ppm. A reasonable result, since you will only see the temperature drift of those 4 or 5 ppm. And they may have selected those switch ICs to use the better ones for the main PWM and the other ones for the helper.

Regards, Dieter

[Kleinstein]

The upper switch is for switch resistance compensation to improve linearity. This is similar to the Fluke 5500.  I would not consider the switch resistance a problem. The extra OPs and switches should compensate for something like 99.x % of the switch resistance, including different temperature dependence at different (but fixed) voltage levels.

For linearity there are likely other weak points:
One is the use of 3 PWM stages and thus a relatively large influence of the middle part. So the trimmer to adjust the influnce of the middle part can be relatively impotant.
Another point is coupling between the PWM stages, if the transitions at some code are in or close to coincidence, this can effect the timing or reference. To reduce this it would take really good supply decoupling and a good layout.
The voltage ranges are a little odd when starting from 7 V, but probably because 2 V and 20 V are common with DMMs, so more or less dictated by demand.
With the switches the speed of switching and maybe charge injection can be a slightly weak points, not so much the resistance.

I don't think the LM308 is happy with a 10 µF load, without some extra series resistance.

The specs look quite optimistic for a LM399 based ref. This would need some burn in and maybe selection to reach this level.

[macaba]

I've built an extremely linear PWM DAC using a similar switch resistance compensation arrangement to IC11/IC8. It reduced the non-linearity by approximately 26x in my case. So with that in mind I'd say the AD7512 looks to be acceptable in this circuit. You could possibly get a slight benefit from ADG333A but the packages don't match and not worth it IMO.

For ripple compensation it is easy to add, e.g. https://www.edn.com/cancel-pwm-dac-ripple-with-analog-subtraction/.

There's nothing special about this circuit - it's just another form of 2nd order RC. Easily proved in ltspice.

[dietert1]

Is this a complaint? I gave you the link, in case you can use it to your advantage. If you can't, that's fine.

Regards, Dieter

[guenthert]

[..]
Linearity and temperature sensitivity
The PWMDAC switches between the reference voltages using high-resistance CMOS switches (AD7512). These switches are about 60 ohms, but that varies with Vds and temperature. This switch is in series with precision wirewound 100 ohm resistors.
[..]

      Perhaps I'm misunderstanding, but to me it looks like the switches are not only in series with 100Ohm resistors, but also with R8, R9 and R20 for a total of some 100kOhm.  The variance in resistance of the switch should then be less of an issue, shouldn't it?

[guenthert]

I've built an extremely linear PWM DAC using a similar switch resistance compensation arrangement to IC11/IC8. It reduced the non-linearity by approximately 26x in my case.

     Umpfh, I might be particularly (or just regularly) dense here, but I don't see how IC11/IC8 (or the corresponding arrangement in the Fluke calibrator) compensates any variation in the resistance of the used analog switches.  To me it rather looks like they attempt to "flatten the curve" when charging/discharging the capacitor (or rather low pass filter element) by adjusting the source given the current potential.  I guess, I need another hint.

[pigrew]

The specs look quite optimistic for a LM399 based ref. This would need some burn in and maybe selection to reach this level.

Yes... they do.

I spent some time trying to measure the output vs a 3458A. (Not so carefully, all in the 22V range of the calibrator). It seems like the low bit definitely needs tweaking. The LSB in this range is about 3.3uV. The other bits are 844uV and 219mV. I don't think it's using delta-sigma, even though it said it was in the manual.

The calibrator's 22V range was calibrated using the DMM at -10V, 0, and 10 V. A few dozen uV of DC offset accumulated (maybe due to noise in the ADC?), but that's only a count or two. The plot with 10uV steps shows periodic 17uV jumps while the max jump should be about 3.3 uV based on the LSB.

Zooming out to 2mV steps, some slight periodicity is visible (219 mV steps), but it isn't bad compared to the LSB.

Looking at full scale, between -10 and 10 V is quite linear. However, strong curvature is seen outside this range. The output is 80 uV (8 counts) low at -22 V. Positive voltages are somewhat better. This seems to be outside the advertised spec limits.

Python code is attached.

[Kleinstein]

The switch resistance compensation works in the following way. Ideally (if all resistors have the right values) the auxiliary switch provides essentially all the current. It sees the slighly higher voltage and than can provide the full current. The main switch would than only correct residual errors, though with the smaller resistance. Not sure if the value for R7 is absolute accurate, considering the switch resistance.

The jumps in the INL curve look a little like the trimmer RV1 is not perfectly set.

The curvature at the larger range can be cause by thermal effects from the amplification - this can give an extra U³ term from power proportional U² and a linear TC of the resistor. Such an effect may also apply to the 100 V range of the 3458 - though I would not expect such a large error.

[MK]

Specifications
LM399?
It's not a LTZ1000A, but is likely good enough for 6.5 digits. I don't have any complaints, though some small tweaks are possible. The Zener current is 930 uA (by my calculation), but maybe it should be increased to the standard 1 mA. As another tweak with negligible effect, I'd power the heater with +/-15V, instead of just the positive rail, so that ground currents are reduced. Adding a series heater series resistor wouldn't be a bad idea, either.

Internally the LM399 runs the zener at approx 0.5mA, any current over that passes through the bypass transistor, there in no need to disturb that section of your calibrator. there are tweaks to the power up circuit using a series resistor and a capacitor across the heater terminals to reduce any start up stress in the die by reducing the max power applied during any start up. The details are in the LM399 spec pdf.

[macaba]

Umpfh, I might be particularly (or just regularly) dense here, but I don't see how IC11/IC8 (or the corresponding arrangement in the Fluke calibrator) compensates any variation in the resistance of the used analog switches.

Kleinstein explained it well. You design the auxiliary switch & R7 combination to supply a nominal 100% of the current. Due to error sources (resistor tolerance, variation of switch resistance over temperature), it ends up supplying 95-105% of the current, and the main switch corrects the value. The much-reduced current in the main switch results in less voltage drop through the switch to a point where the resistance mismatch becomes insignificant.

[guenthert]

     Umpfh, I might be particularly (or just regularly) dense here, but I don't see how IC11/IC8 (or the corresponding arrangement in the Fluke calibrator) compensates any variation in the resistance of the used analog switches.  To me it rather looks like they attempt to "flatten the curve" when charging/discharging the capacitor (or rather low pass filter element) by adjusting the source given the current potential.  I guess, I need another hint.

      Well, for that to work the polarities of the compensation circuity would have to be reversed.  I was considering whether the schematic was wrong, but it's likewise on Fluke's 5700 PWM DAC board; there oddly, only the positive leg ("series linearity control") is pulsed synchronously to the MSB PWM and uses a fixed source, while the negative leg ("shunt linearity control") is permanently injected, but its source depends on the present output potential.
     So that seems to be intended to counteract the non-linearity of the on-resistance of the analog switch as it depends on the voltage across it.  I can only assume that the non-linearity from the charging of the capacity is counteracted in software, as that's at least well known and fixed.

[Kleinstein]

Th epolarity of the auxiliary votlages looks OK if the switches are operated in parallel. The positive auxiliary votlage gets somewhat larger than the reference. The ideal is to ignore the AC part at thin filter input. So when the postive ref. is active the current through R8 is the same as the current through R16, R18 and R7. So there is essentially no current through the main switch. The analog way applies to the negative side.  One could argue that R18 and R19 should be larger by the switch resistance to improve the balance.

The Fluke5700 circuit is quite complicated as it includes extra gain and not just a buffer. As far as I understood it, the principle is still the same but in a really confusing way.

[dietert1]

When you search the web for "precision pwm" you find all kinds of contributions, but little useful for this type of calibration PWM. As far as i understand other points beyond the linearity of the switches are
- Timing:
Many of these CMOS switches have delay times of 30 or 60 nsec. A 1 nsec asymmetry between the upper and the lower switch makes a 1 ppm shift at 1 KHz PWM frequency. By the way: At what frequency are they operating the PWM in this calibrator? Little is known about temperature dependency of those delay times.
- Static asymmetry:
With a 40.2 KOhm PWM resistor a 1 Ohm asymmetry between upper and lower switch shifts output by about 10 ppm. At least that's what i found when making a PWM for a 7 -> 10 V gain stage. As we know channel resistance rises with temperature, for example by 20 % between 20 and 50 °C. At 60 Ohm channel resistance this can give you a temperature dependency of the PWM output. I think the redundant switch arrangement is meant to reduce this effect.

I would check the accuracy of a PWM with a modern digital scope. At some Gsamples/sec you can determine the digital PWM value at the input of the analog switch with good accuracy. And then measure what is the ratio on the analog side using a precision DVM. The agreement should be better than some ppms. Only then you may assume that the DAC precision will be a ppm or so.

Regards, Dieter

[Kleinstein]

The extra switches are there to reduce the effect of changing switch resistance, if done right it can be quite effective. Even with the switch resistance missing at R18/R19, it may compensate 98-99%.  In a first approximation the change is similar for the low side and high side, only the difference matters.
For the switch delays it is also about the change in the delay, not so much die absolute values. The DAC gain is included in the calibration measurement.

I would expect a PWM frequency somewhere from 200 Hz to 1 kHz.

The temperature dependence of the delays is indeed a possible weakness.

[dietert1]

Sorry, pigrew already wrote "PWMDAC (at about 4 KHz)". That means a 1 nsec shift can cause a 4 ppm output voltage shift.

Regards, Dieter

[pigrew]

Sorry, pigrew already wrote "PWMDAC (at about 4 KHz)". That means a 1 nsec shift can cause a 4 ppm output voltage shift.

Regards, Dieter

I apologize, but I've started to find a few errors in my first post. Once I have a more clear idea of what's happening, I'll update the first post, and make another one.

PWM Frequency

The PWM frequency is actually 244 Hz (as measured by oscilloscope on all three switch outputs). It's driven by a 1 MHz oscillator mounted on the analog board.

Non-linearities

I'm also starting to think that the non-linearities at around +20 and -20 V are caused by non-linearities in the amplifier and not the PWMDAC. I'm trying to wrap my head around the TR200/TR201 output stage, as it isn't something I've seen before. Simulation is suggesting that they are saturating when the output is programmed to be -20V.

EMI

Also, the mysterious 900 kHz signal seems to be present even without anything connected to the Rigol DSA823E spectrum analyzer. Connecting to the calibrator with a BNC to banana adapter is picking up a ton of EMI (with calibrator turned off). I have seen occasional oscillations (oscillation frequency changes when amplifier is switched between gain ranges), but it isn't so repeatable. The digital I/O between the MCU and analog board seems silent except when the output is reprogramed, so its lack of a GND wire wouldn't cause EMI except when changing output.

Stability
LTSpice simulation shows a phase margin of 5 degrees with values actually populated on the analog board. Using values from the schematic, the phase margin increases to 43 degrees. Based on that, I've modified the amplifier compensation to have the schematic-specified values.

[Kleinstein]

The output stage is really a little odd looking - like from times when transistors were very expensive.  TR200/201 are in base configuration. The compensation looks awfully complicated, especially considering that a simulation like spice was not that common at that time.

I don't think the transistor part is responsible for INL near the +-20 V ends. The transistor part is inside the control loop of the OP - so the INL from this part is expected to be more abrupt (e.g. close to hard clipping at some 20.x V). So the saturation may be wanted to get protection from spikes to exceed the 20 V by much.
Over the large range the INL looks more like a U³ term as expected from thermal effects. The indicated S102 resistors should be good though.

[guenthert]

      In Fluke's 5700 the OpAmp follower right after the filter has its supply boosted to minimize non-linearity caused by finite common mode rejection.  The Xitron 2000M skips that, so perhaps there's the source of the observed non-linearity.

[Kleinstein]

The linearit of the buffer could be an issue. The CMRR is quite good, but the open loop gain is not very high. One may be able to measure the linearit also directly at the PWM DAC, so before the output amplifier.

[pigrew]

Based on LTSpice, the amp's gain (22V range) goes down by 37 ppm when trying to output -22 V.

Also contributing to the non linearity plot is that it the bandwidth is low enough that it's very slow for the output to settle. I may not have delayed enough before the first point when doing my measurement (I only waited 90 seconds).The manual says "Settling Time: ---------- Less than 2 seconds to within 10ppm of change + 0.25μV", so I figured 90 seconds would be sufficient. I should have measured the hysteresis.

(Added later: Thinking for another minute, I realized the plots are the AC gain, so would need to be integrated over voltage to figure what the actual DC offset would be at the output. But, they do show that the amplifier is struggling at the edges of its range; DC op point simulation has been attached.)

[Kleinstein]

Using the AC simulation to judge on settling can be misleading. Even if the AC gain is reduced quite a bit, the settling to a jump can be quite a bit better than one may expect from the simple picture.  For the settling after a jump it is more like looking for the time constants and than assume an exponential decay for the settling.
For setling I would mainly consider the filter circuit at the DAC - the amplifier is surprisingly slow, but still faster than the filter.
 
For the very last ppm range one may have to check if there are capacitors with high DA - these can in some cases cause an nasty surprisingly slow extra contribution.
For settling time to the 10 ppm level the type of capacitor (e.g PP or polyester) can make a difference.

[MegaVolt]

Tell me where did you get the schemes? I can't find any documents

[pigrew]

Tell me where did you get the schemes? I can't find any documents

I asked Xitron for a copy of the service manual (via email). The service manual had very close to correct schematics. I had reverse-engineered the power supply board before I thought to ask for schematics.