General Purpose Power Supply Design

[fmaimon]

A little update. I did a spreadsheet to calculate the PWM value for each voltage, then I manually set this value to the tiny2313 and recorded the output value. I've used the 1V and 9V output values to calibrate the sheet and recorded the output value read on my 34410A (10V range, 10 NPLC). The results are pretty good!

 

	Voltage setting	34410A
Voltage	PWM value	Output
0,5	807.138	0,49998
1,0	1.657.807	1,00002
2,0	3.359.144	2,00001
3,0	5.060.481	3,00002
4,0	6.761.818	3,99999
5,0	8.463.156	4,99999
6,0	10.164.493	6,00000
7,0	11.865.830	7,00002
8,0	13.567.167	8,00003
9,0	15.268.504	9,00003
9,5	16.119.173	9,49872

[BravoV]

fmaimon, thanks for the update, thats even good enough for me to build that module alone as an adjustable reference  , any chance you could post the overall circuit and code, please ?

[eevblogfan]

hey

WoW , nice results ! , good job !!

[fmaimon]

The source code is attached. It's pretty much Richard's (amspire) code with a little bit of crude uart control.

I haven't really drawn any schematic but the circuit is pretty straightforward. The only thing I didn't say in the other post is that I've used a 3 stage 47k/220nF filter.

[amspire]

Felipe,

Great results - matched my experience. Did your development board use a crystal or ceramic resonator for the 20Mhz clock? Also does the Tiny2313 have a low power and full swing clock oscillator mode (like the atmega's) or just a single power clock mode?

The reason I ask is that I found I had to use the full swing clock mode on the Atmega328P. In the low power clock mode, the jitter on the clock edges was enough to cause a noticeable instability on the output. In the full swing mode, the jitter was less then 1nS, and the PWM results were fabulous. It appeared that in the low power mode, the XTAL1 input was sensitive enough to be modulated by switching noise from the micro outputs.

Richard.

[fmaimon]

Felipe,

Great results - matched my experience. Did your development board use a crystal or ceramic resonator for the 20Mhz clock? Also does the Tiny2313 have a low power and full swing clock oscillator mode (like the atmega's) or just a single power clock mode?

The reason I ask is that I found I had to use the full swing clock mode on the Atmega328P. In the low power clock mode, the jitter on the clock edges was enough to cause a noticeable instability on the output. In the full swing mode, the jitter was less then 1nS, and the PWM results were fabulous. It appeared that in the low power mode, the XTAL1 input was sensitive enough to be modulated by switching noise from the micro outputs.

Richard.

Richard,

Actually it was a bit better. I was using the 10M input of my 34410A so, together with the 3 47k resistors, it had a little voltage drop. Once I've put the input in Hi-Z mode, there were no measurable diference between the reference and the output in the highest output value. It's incredible that even with this "voltage divider" it could maintain a great linearity.

I've used a crystal connected to the STK500's crystal oscillator. It is an external clock signal. I've selected the clock-out fuse of the t2313 and the clock seems stable, but I really don't know how to check the jitter. How do you do that?




Reading the datasheet, looks like it doesn't have a specific full swing mode for the crystal oscillator. I can check later.

[amspire]

I've used a crystal connected to the STK500's crystal oscillator. It is an external clock signal. I've selected the clock-out fuse of the t2313 and the clock seems stable, but I really don't know how to check the jitter. How do you do that?

An external clock oscillator is probably an optimum solution.

You measure jitter with an oscilloscope.

Look at the triggered edge first - is the edge very narrow with no jitter evident at the fastest timebase speed? Look at the thickness of the edge and estimate how much jitter you would be able to see. If the scope can do 2nS/div, and the edge variations are less then 0.1 divisions wide, then you can hopefully see jitter down to 0.2nS. For a digital scope, you probably want to use equivalent sampling to increase the sampling rate above the A/D converter's sampling speed.

To see the jitter, connect the scope to the PWM output. Lets say the PWM output goes high at zero count, and low when the count is reached. Trigger the scope off the positive edge and use delayed sweep, or a big sampling buffer so you can look at the next positive PWM edge at the scopes fastest sweep rate. You want to see the variations in the edge timing. For an analog scope, turn the brightness right up, and look at the thickness of the positive edge. The more jitter, the thicker it will be.

For a digital or storage scope, get the scope to accumulate all the edges, and look at the variations. you might leave it running for several minutes to see if there are any spurious edge errors. You may have to juggle a bit to get the best out of a cheap digital scope - they often do not have a genuine delayed trigger.

If you have a stable squarewave oscillator, you can use it to check the validity of the jitter results.

Richard.

[fmaimon]

You measure jitter with an oscilloscope.

LOL. 
I knew that one... 

Look at the triggered edge first - is the edge very narrow with no jitter evident at the fastest timebase speed? Look at the thickness of the edge and estimate how much jitter you would be able to see. If the scope can do 2nS/div, and the edge variations are less then 0.1 divisions wide, then you can hopefully see jitter down to 0.2nS. For a digital scope, you probably want to use equivalent sampling to increase the sampling rate above the A/D converter's sampling speed.

To see the jitter, connect the scope to the PWM output. Lets say the PWM output goes high at zero count, and low when the count is reached. Trigger the scope off the positive edge and use delayed sweep, or a big sampling buffer so you can look at the next positive PWM edge at the scopes fastest sweep rate. You want to see the variations in the edge timing. For an analog scope, turn the brightness right up, and look at the thickness of the positive edge. The more jitter, the thicker it will be.

For a digital or storage scope, get the scope to accumulate all the edges, and look at the variations. you might leave it running for several minutes to see if there are any spurious edge errors. You may have to juggle a bit to get the best out of a cheap digital scope - they often do not have a genuine delayed trigger.

If you have a stable squarewave oscillator, you can use it to check the validity of the jitter results.

Using the STK500 external oscillator, I seem to be seeing about 1 ns jitter. I have a TDS2014B and I'm moving the horizontal position 500n - 1us to the left of the screen, so I should be seeing the waveform this far from the trigger. It doesn't have a big buffer, so it was the only way I could think of. I also used infinite persistence.

This weekend I'll etch a board to try the RC oscillator of a tiny45, to try it's PLL and it's crystal oscillator. I'll check the jitter.

As I'm thinking of using the t45 and it's PLL, I did optimise the interrupt code (pretty much took the compiler output code and tweaked it), and came out with this:

Code: [Select]

#include <avr/io.h>

.extern Out_Val
.extern pwm_accum

/*
    Interrupt routine for timer 0
    Its the 24 bit delta sigma converter,
    generating the PWM output.
*/
.global TIMER0_OVF_vect
TIMER0_OVF_vect:
    push r20
    in     r20, _SFR_IO_ADDR(SREG)
    push r20
    push r21

    // loads third byte from pwm_accum and stores on OCR0A
    lds  r21, pwm_accum + 2
    out     _SFR_IO_ADDR(OCR0A), r21

    // Add the original intended value to the error in the accumulator for the next PWM cycle
    lds  r21, pwm_accum + 0
    lds  r20, Out_Val + 0
    add  r21, r20
    sts  pwm_accum + 0, r21
   
    lds  r21, pwm_accum + 1
    lds  r20, Out_Val + 1
    adc  r21, r20
    sts  pwm_accum + 1, r21
   
    // Clears the top byte of pwm_accum and add the original value
    andi r21, 0
    lds  r20, Out_Val + 2
    adc  r21, r20
    sts  pwm_accum + 2, r21

    pop  r21
    pop  r20
    out  _SFR_IO_ADDR(SREG), r20
    pop  r20
    reti

Just save this code as "isr.S" or similar and add to you project on AVR Studio. Beware of the uppercase "S". Without it, it won't "compile". It's 42 cycles long so, even with a 64 MHz timer clock and a 16 MHz main clock, it still have 22 cycles to spare. I've tried this code and it seems to work.

As I'm not an assembler guru, someone may still think on a way of making it smaller and faster.

Felipe

[amspire]

This weekend I'll etch a board to try the RC oscillator of a tiny45, to try it's PLL and it's crystal oscillator. I'll check the jitter.

Random Jitter is probably not a huge concern. Modulation caused by the changing micro supply current, and switching outputs is a big concern. Say that every time a particular output pin goes high, there is an extra 0.01V supply drop on the micro chip, and this causes the on-board oscillator to run at a slightly different speed. That will wreck the accuracy. Crystal oscillators and ceramic resonators will be far less susceptible to this problem then an on-board RC oscillator.

Also the on-board RC oscillator is not really accurate enough for reliable RS232 type serial protocol which is a big pain. Not sure it is worth it.

As I'm thinking of using the t45 and it's PLL, I did optimize the interrupt code (pretty much took the compiler output code and tweaked it), and came out with this:

I am dubious about the benefits. If the PLL is not extremely stable, then it will make it even worse then no PLL at all. It is worth a test. The thing I liked about the PLL running at the crystal clock speed is that the interrupt period can easily cope with managing 2 PWM DACs, plus leave a fair margin for other interrupts - as long as all other interrupts re-enable the PLL interrupts on entry. When you start using the tiny45 PLL for the PWM clock, the processing time for the interrupt gets extremely tight as you cannot afford to miss even one interrupt if you want accuracy. It can work if the tiny is dedicated to the PWM task, and it doesn't have to do much else.

Richard.

Edit:  I am talking about errors that will probably be in the 4th or 5th decimal place. For a system that needs at best a 0.1% error, then there is a good chance an internal RC oscillator, and the tiny45's PLL will probably be fine. So the tests are definitely worth while, even if it just to learn how good the internal RC oscillator is, and how good the PLL is. I am very interested to find this out.

[ali_asadzadeh]

Dear Richard
Thanks for sharing all the good stuff.
The good news is that Cortex-m0 parts with 16bit ADC and low price(bellow 1USD) are on the way. (freescale kinetis L0 MCU's)

I think they sound like good candidates for digital power supplies, and I think TL431 is a good candidate for reference part.
Can we achieve 16bit accuracy and stability with a low cost reference such as TL431?
How can we go to 40v 10A region with lowest cost?
As I see one of your main concerns is cost, I think it should be your main concern!
Again Thanks for all your great work.

[amspire]

Dear Richard
Thanks for sharing all the good stuff.
The good news is that Cortex-m0 parts with 16bit ADC and low price(bellow 1USD) are on the way. (freescale kinetis L0 MCU's)

I think they sound like good candidates for digital power supplies, and I think TL431 is a good candidate for reference part.
Can we achieve 16bit accuracy and stability with a low cost reference such as TL431?

No but I am considering the TL431 for the power supply as it is a great reference for price, multiple sources and for a low supply voltage. The 431 means I can, say, have the micro running at 3.3V from a low dropout regulator, and have a TL431 reference running at 3.0V powered from the micro's 3.3V supply. This means I can let the input voltage for the general purpose supply go from about 3.5V to 30V.

I never tried to make the general purpose supply precision, but I did want it to have much better then 1mV adjustability. If you need an accurate output, you can always let the supply stabilize thermally, and then adjust the output with an external DVM, as long as the supply has the resolution to let you adjust it. I do not see any huge need for a supply to have better then about a 0.25% accuracy, as long as you have an accurate DVM for the few times you need a precision output.

For a precision reference, you have to start looking at the references with the temp coefficients in the low ppm, higher regulated supplies for the reference supply, and precision  opamps and precision resistors/voltage dividers if you need to amplify or attenuate the reference output.
How can we go to 40v 10A region with lowest cost?

As I see one of your main concerns is cost, I think it should be your main concern!
Again Thanks for all your great work.

The ideal has always been a basic 1A supply module that could cost less then $10. I do have a plan for extending the supply to any voltage you like (up to hundreds of volts), and the plan for more then 1A has always been to parallel the basic module - all under the control of a central display/control board.

Richard.

[fmaimon]

Here is the schematic I'm planning to build as a test bench. What do you think?

[amspire]

Here is the schematic I'm planning to build as a test bench. What do you think?

It depends what you want. The LT1013 is not a suitable replacement for the LM324 at all for the power supply itself - it cannot handle the 32V on the inputs that the LM324 can handle, so in spite any claims from Linear Technology, it is not fully compatible.

If you are making just the reference, there are a few changes I would do. I would probably go for autozero opamps, and definitely a very low bias current MOSFET or JFET opamp. there are great cheap 5.5V auto zero opamps, but there are less available when you look at 30V opamps.

The main change I can suggest is to change the output amp to something like this:



It allows you to do a few things. It means the output can go above 10V, and you can also set the zero volts out (particularly if you can let the opamp have a 0.5V negative supply). It is always great if a reference can go from slightly below zero on the output to about 10% over-voltage. It means you can see what reference voltage output will get a meter reading 10.00000V. You will find if you can go 10% over voltage, you will use the feature regularly.

If you have set the 10V REF01 to be accurate from a Geller reference (or equivalent), then you can use software calibration to set the 10V output of the programmable reference (you just fine what voltage number gives zero difference between the 10V ref out and the programmable ref out), and you find out what voltage number gives 0V out. Once you have these two calibration numbers, you can calculate the number for all other voltages.

Also, to get the best accuracy from the reference, you want to pre-regulate the 15V that supplies the REF01. Also, you want it battery powered. Possible a pair of rechargeable 9V batteries will do the job. Eliminating AC supplies/power packs from a reference eliminates some of the potential error sources.

Richard.

[fmaimon]

It depends what you want. The LT1013 is not a suitable replacement for the LM324 at all for the power supply itself - it cannot handle the 32V on the inputs that the LM324 can handle, so in spite any claims from Linear Technology, it is not fully compatible.

I'm no replacing the LM324. I'm just using a precision opamp (low offset drift) to buffer the output. As the

If you are making just the reference, there are a few changes I would do. I would probably go for autozero opamps, and definitely a very low bias current MOSFET or JFET opamp. there are great cheap 5.5V auto zero opamps, but there are less available when you look at 30V opamps.

I didn't use autozero opamps in this design as I don't have any right now. I'm just design with what I have at hand.

The main change I can suggest is to change the output amp to something like this:

I'm assuming that the opamps input in this circuit are inverted and the bottom of the 91K resistor is ground, right? What is the 10M resistor doing?

It allows you to do a few things. It means the output can go above 10V, and you can also set the zero volts out (particularly if you can let the opamp have a 0.5V negative supply). It is always great if a reference can go from slightly below zero on the output to about 10% over-voltage. It means you can see what reference voltage output will get a meter reading 10.00000V. You will find if you can go 10% over voltage, you will use the feature regularly.

Right now I don't want to put things that have tempco. Right now it is just the REF01 and the offset drift of the LT1013. One solution is using two resistors in the same package. This one seems a good option, with 5 ppm/C tracking between the resistors and is cheap, although I never heard of this manufacturer. Here is the datasheet: http://datasheet.octopart.com/RM3216A-103/903-PBVW10-Susumu-datasheet-119922.pdf

If you have set the 10V REF01 to be accurate from a Geller reference (or equivalent), then you can use software calibration to set the 10V output of the programmable reference (you just fine what voltage number gives zero difference between the 10V ref out and the programmable ref out), and you find out what voltage number gives 0V out. Once you have these two calibration numbers, you can calculate the number for all other voltages.

That is the idea. Instead of a Geller reference, I'll just use my pair of trusty meters, an Agilent 34410A, about 1 year old, and a Fluke 8840A, lots of years old and calibrated this january. Although the 8840A is only a 5 1/2 meter, the readings on both are pretty much the same.

Also, to get the best accuracy from the reference, you want to pre-regulate the 15V that supplies the REF01. Also, you want it battery powered. Possible a pair of rechargeable 9V batteries will do the job. Eliminating AC supplies/power packs from a reference eliminates some of the potential error sources.

Right now I'm just keeping it simple, as this is the first prototype. The power from the circuit will be my bench power supply. Later, when integrating everything together with your power supply design, I'll do it properly.

[amspire]

I'm assuming that the opamps input in this circuit are inverted and the bottom of the 91K resistor is ground, right? What is the 10M resistor doing?

Yes, I got the opamp connections the wrong way round. I meant to have the -ve input to the feedback resistors.
The 10M resistor is offsetting the 0V point, but it can be much higher (to reduce temp coeff. issues). This allows you to actually calibrate the 0V point.

It allows you to do a few things. It means the output can go above 10V, and you can also set the zero volts out (particularly if you can let the opamp have a 0.5V negative supply). It is always great if a reference can go from slightly below zero on the output to about 10% over-voltage. It means you can see what reference voltage output will get a meter reading 10.00000V. You will find if you can go 10% over voltage, you will use the feature regularly.

Right now I don't want to put things that have tempco. Right now it is just the REF01 and the offset drift of the LT1013. One solution is using two resistors in the same package. This one seems a good option, with 5 ppm/C tracking between the resistors and is cheap, although I never heard of this manufacturer. Here is the datasheet: http://datasheet.octopart.com/RM3216A-103/903-PBVW10-Susumu-datasheet-119922.pdf

The divider solution I like is the Linear Technology ones. The ratio tracks to 0.2ppm/C, and the price is reasonable.
http://parametric.linear.com/precision_resistor_network

[HackedFridgeMagnet]

Hi fmaimon

Here is the schematic I'm planning to build as a test bench. What do you think?

One thing to check, don't you want to tie "not reset" high via a resistor? For when the isp is not connected.

[fmaimon]

Hi fmaimon

Here is the schematic I'm planning to build as a test bench. What do you think?

One thing to check, don't you want to tie "not reset" high via a resistor? For when the isp is not connected.

All AVRs have an internal pull-up resistor on reset. Usually, it's only really necessary to put another lower value resistor in high noise enviroments. This is not the case here.

[fmaimon]

Yes, I got the opamp connections the wrong way round. I meant to have the -ve input to the feedback resistors.
The 10M resistor is offsetting the 0V point, but it can be much higher (to reduce temp coeff. issues). This allows you to actually calibrate the 0V point.

...

The divider solution I like is the Linear Technology ones. The ratio tracks to 0.2ppm/C, and the price is reasonable.
http://parametric.linear.com/precision_resistor_network

Nice resistors! Digikey's search don't have a track ratio option, so I've missed it. As the LT5400-3 has 4 resistors (2x 10k and 2x 100k), I've changed your opamp design a little bit.



The LT5400-3 are R1, R2, R3 and R5. R4 is just an output load.

Now I don't have to worry about the offset resistor tempco and the output can go from -1V to 11V, making possible to calibrate 0V and 10V points.

Felipe

[fmaimon]

One more thing, regarding the opamp. Is there any problems in using one like the AD8638? It has a higher offset, but as the output voltage will be calibrated, it shouldn't make a difference.

In quantities of 25, the SMD version of the LTC1050 is $3.60 and the AD8638 is $2,49. The LTC1050 have a cheaper dip option at $2,87.

[ali_asadzadeh]

Dear Richard
Does OP07 considered a good choice for 40V version? and I want to use ADR421 for reference, which resistors do you suggest for dividing the output voltage down to 2.5V?

[amspire]

Dear Richard
Does OP07 considered a good choice for 40V version? and I want to use ADR421 for reference, which resistors do you suggest for dividing the output voltage down to 2.5V?

Is this for the supply, or a precision reference?

The OP07 and its close relatives are good, but I would probably go in a different direction the reference if you are chasing precision. I would probably use something like a Microchip MCP6V27 dual auto zero opamp with microvolt offset and extremely low input current. I would use a NPN transistor or N channel mosfet on the output with a collector/drain resistor to the positive supply to boost the voltage level to 40V or more.

I did discuss something similar in

https://www.eevblog.com/forum/projects-designs-and-technical-stuff/100vdc-reference-circuit/
https://www.eevblog.com/forum/general-chat/ltc1052-fifth-i-have-destroyed-whilst-prototyping-any-idea-why-they-are-dying/

The thing to watch out for in this configuration is that it is easy for the divider from the output to drive the inverting input above the opamp's supply rail causing the opamp to latch up. Some kind of protection is needed, particularly on startup.

For the divider, I would stick with the Linear Technology dividers that I mentioned before. A package with 2x10K and 2x100K can easily be used for a very stable 2.5, 5, 10, 20 or 40  times divider. If you only need around 0.1% accuracy, then 25ppm/C metal film resistors are fine, and a lot cheaper.

I do have plans to redesign my supply with a floating regulator that can regulate way beyond the 30/40V limits of common opamps - it would let the regulator go to 100's of volts - and that would mean a different arrangement for the reference output. It would be in the form of a current source, with a resistor to ground to generate the required voltage.

Richard.

[ali_asadzadeh]

Thanks for your reply.
I want to use a low voltage reference so that I could use that for Kinetis MCU reference, It has 16bit ADC, in that way I can read back the output voltage. And the OP07 is regulating the 40v output.
Do you think it's a good idea? Is there a better way?
Also I have another question,
Suppose that I want to have dual output voltages, is there a kind of isolated voltage references, I mean is there a cheap way of using voltage of the reference IC for other parts of the circuits, I mean something like opto couplers, or something like Analog devices  RS485 transceivers (ADM2682E)  that they generate an isolated supply voltage on the chip, is there any option for using the reference output  with isolation? Or I should use another Reference chip?

[ali_asadzadeh]

There is one more question.
You have used a 3 stage RC filter after the PWM, so what happens to capacitor and resistors temperature coefficient in there? I mean there is no feedback there, and you have used normal R and C for there. As you know if temperature changes the Q factor of filters will change and will change filter response, does this affect performance?

[amspire]

There is one more question.
You have used a 3 stage RC filter after the PWM, so what happens to capacitor and resistors temperature coefficient in there? I mean there is no feedback there, and you have used normal R and C for there. As you know if temperature changes the Q factor of filters will change and will change filter response, does this affect performance?

The temperature coefficient  and drift of the resistors has zero effect on accuracy - you could use carbon film resistors and the result would be just as stable.

As long as the temperatures of capacitors are allowed to stabilize, the coefficient has zero effect. If a capacitor was in the process of warming and the capacitor has a significant temperature coefficient, it could vary the voltage slightly. The capacitors will generate no internal heat, but it would be a good idea not to place any of the capacitors near a hot component.

The source of the error would be this. If the last capacitor was increasing temperature by 10 degrees per second, and it had a positive temp coefficient of 1000ppm/C, then in 1/10th of a second, it could increase in value by 100ppm. This could cause the output voltage to be low by perhaps 10 to 20ppm for the period the capacitor is changing temperature. Not a big error but an error. The most important factor in choosing a capacitor is extremely low leakage, but second would be to pick the lowest convenient temp coefficient.  With all precision test gear, if you want the most accurate performance, you let the equipment warm up for perhaps 30 minutes, and if you do, then the temperature coefficient does not matter.

Long term stability of the capacitor does not affect the accuracy at all.

Richard.

[amspire]

Thanks for your reply.
I want to use a low voltage reference so that I could use that for Kinetis MCU reference, It has 16bit ADC, in that way I can read back the output voltage. And the OP07 is regulating the 40v output.
Do you think it's a good idea? Is there a better way?

If you are going to change to a different family of processors, you will have to test the PWM performance of the new processor. The Atmega's are great as they are able to get the jitter on the PWM edges down below 1nS. That is the secret of the linearity and stability. The Kinetis may be great too - I would not know. I also do not know if the PWM of the Kinetis chip has the right options to implement the PWM properly.

I posted my view a short while ago on how I would handle high voltage reference outputs. I prefer something like a MCP6V27 and a transistor on the output to get to 40V or more.

My GP supply needs the LM324 as the supply regulator IC. No substitutions possible.

Also I have another question,
Suppose that I want to have dual output voltages, is there a kind of isolated voltage references, I mean is there a cheap way of using voltage of the reference IC for other parts of the circuits, I mean something like opto couplers, or something like Analog devices  RS485 transceivers (ADM2682E)  that they generate an isolated supply voltage on the chip, is there any option for using the reference output  with isolation? Or I should use another Reference chip?

I would use a separate micro, ADC (if you want better then the micro A/D)  and reference IC for each supply. If you then have an intelligent front panel and/or USB, use another micro again for the front panel which controls the power module IC's via optocouplers.

If you coupled the PWM via an optocoupler to a floating supply, it would work, but the accuracy and stability would decrease. You may get down to 0.1% accuracy, but you may have to settle for 1% accuracy. Still OK for a basic supply. I found that when I used a 4000 series cmos IC as the PWM output, the results were disappointing due to the slow edge speed, and the Optocoupler speed would have a similar effect.

Richard