A modern DC/DC converter design capable of delivering current in excess of 40mA at 170V
Already have an account? Log in.
To make the experience fit your profile, pick a username and tell us what interests you.
These pages describe in great details how I built a reliable and efficient high voltage nixie power supply -showcased above-; from the idea to selling it!
Be sure to read project logs!
Between all prototypes, components, assembly and packaging material, I have spent about US$2854 in this project. Pocket change if I were a corporation, but a hefty sum for an individual! The breakdown is as following:
| Original Prototype (JLCPCB) | $17.90 |
| More prototype experiments (JLCPCB) | $18.63 |
| Components for Prototypes (Mouser) | $146.98 |
| Assembly Prototype (PCBWay) | $286 |
| Production Run (PCBWay) | $2335 |
| Packaging Material (local supplier) | $50 |
| Grand Total | $2854.51 |
The power supply is available for sale at two different places: eBay and Tindie.
Tindie comes with about 9% fees, while eBay is at 10%. In addition to that, the actual cost of shipping is SGD7.20 for an international registered mail, or about US$5.40. The real shipping cost is "hidden" in the product cost simply because I believe people are put off by shipping fees -myself included.
With all these considerations, you can calculate that there's about $6.5 profit per board; under the assumption that my time is free, that there are no defective boards, and that there will be no refunds or lost packages.
At the current rate of 4 boards per week, it would take 34 weeks (9 months) just to get my money back. At this breaking-even point I would have sold 137 power supplies.
My morning routine has changed a bit. Nowadays, before checking in at the office, I do a quick detour by the post office to send a power supply somewhere across the world. It is truly an amazing experience to see that people are spending their hard earned money in a product that you designed from start to finish. This is really gratifying and I am thankful for this.
In reality though, considering how involved this process is, it is absolutely not worth it if you're trying to make some money out of it. The volume of sales you'd have to achieve before getting some returns would have to be at least an order of magnitude higher than what it currently is. You really have to experience it yourself to realize how razor thin the margins on electronics are.
It is, however, a very good experience. I never intended to make money with the project, just fund my hobby at best. On this front, the objective will be attained eventually. I will probably re-iterate the experience, but on a more affordable product that doesn't require exotic components. As of today, I am already working on a nixie tube driver that will be released in April 2019.
Production run is all about making your product the most cost effective. In the world of electronics, volume is king. As I am an independent designer with limited resources, I unfortunately cannot order 1000 pieces upfront. Nixie enthusiasts is a niche market anyway so a 1000 units would take years to sell. I decided to go with 200 pieces.
200 stems from the fact that the most expensive component on the power supply -the power transformer- comes in full reel of 200. Ordering a complete reel of components is typically the most cost effective.
On a turnkey basis, PCBWay quoted US$2270 for 200 boards. That is $11 per board and it doesn’t even include the cost of assembly and PCBs! Needless to say, this was way more than I first anticipated.
The biggest offender, again, is the Coilcraft transformer, at a quoted US$4.40 a piece, despite the volume. For reference, Mouser quotes US$2.84 a piece for a full reel.
I did the sensible thing and decided to source most of the components myself, but first of all I called the Chinese Coilcraft representative to explain my situation; who sent me over to a local counterpart in Singapore. This is how this conversation went:
And so I ordered the complete reel (and an additional 2 pieces to allow for mishaps at the assembly line) from Coilcraft.com.
To that they added $26.77 worth of export US tariffs and $48(!) of shipping for a grand total of US$610.06. Once the parcel reached China, I was slapped with 25% import tariff thanks to the Sino-American economic war; which added US$147 ($140 + $7 paypal fees paid to PCBWay). Total of the operation: US$757.06 or US$3.79 a piece. Still cheaper than Turnkey service but certainly a lot more than what I thought.
Moral of the story: don’t work with Coilcraft if your assembly line is in China. Addendum: I used to love all their quality components; I will now go out of my way to avoid them. Rant over.
Since I had already ordered 10 boards with the exact same Gerber file, this time production went extremely smoothly. In line with what happened the first time, the PCBWay representative will order the production of a single board (or panel in that case), ask me to confirm everything looks good, and from that point it’ll only take a day to do the rest of the production run.
To give you a sense of timeline, from the moment they acknowledged receiving all the components to Fedex picking up the shipment, only 5 working days had passed. The lost time is mostly in the logistic of sourcing the components.
Fedex being Fedex, they claimed “Office closed or customer not available” on the day of the scheduled delivery, at 5:40pm, when there is a receptionist and I was there waiting. No one ever turned up on the day of the delivery. And the update was “Collect at the Fedex location”. You pay so much for courier delivery and that’s how they treat you. After discussing with an agent and threatening that I could pull the CCTV clearly showing no one ever came; they rescheduled the delivery on the next day.
After all of this, I am now the proud owner of 200 nixie power supplies.
Once I received the first professionally assembled boards, there was one last thing to do to confirm the design behaves just as expected: testing it!
I measured the efficiency of the power supply using a rig made out of a female header and a bunch of 30k resistors. That way I can measure loads at roughly 5mA intervals (170V / 30K = 5.7mA per resistor). The good thing about working with high voltage/small current is that you can get away with really small wires. The big chunk of metal used here is more than capable of handling that load.
In addition to that, measurements were made using 3 different source voltages: 5V, 9V and 12V. It would technically be possibly to power the supply with 4.7V lipo but that would require extensive battery protection mechanisms. For that particular reason I decided to not include this test as it might encourage unsafe operations.
Without further ado, here are the results:
At 5V the power supply is pretty impressive. Having over 80% efficiency on this ~34 boost ratio is an amazing result! Past 20mA it quickly drops though, showing the limits of the design at such low voltage. That being said, this makes a nixie clock driven by a single 5V wall wart completely possible! Very exciting!
9 and 12V are showing similar curves, and the power supply can output well over 40mA@170V using these. On very low current (~5mA) the 12V input shows a lackluster 71% efficiency but it rapidly increases to 85%+ efficiency as hte load increases.
I pushed the power supply all the way past 50mA but past that it’s simply just an absurd amount of current to power tubes. It simply means the power supply operates well within its capabilities and will be extremely reliable. Those tests are a complete pass!
Before sending the board for assembly, I added some final touches in the form of a Revision C:
The changes compared to version B are very subtle, but you can note:
I didn’t want to order one last prototype. At this stage I was over 3 months in this project, and a few hundreds dollars out of pocket. The changes were so minor I decided to roll with it all the way to the assembly.
PCBWay has been heavily advertising their “US$88 for 10” assembly and it seems like this is an absolutely unbeatable price. So I decided to trust them and sent in as per usual my Gerber files for the PCBs. Together with it, they of course ask for a BoM (Bill of Material).
The funny thing is there is no industry standard for BoMs. It’s just an Excel spreadsheet with the list of components, and each fab will have different requierements. At PCBWay they need:
I tried to be as thorough as possible, and also provided them with a list of alternatives when possible.
PCB fabs will produce your board down to the exact spec you give them, but the thing is a lot of components absolutely don’t matter. For instance in this case the automotive version of a 140k resistor will work just as fine as the regular part but for some odd reason your fab might only have automotive parts available at the moment.
In any case, once you’ve sent the BOM the fab will get back to you with prices and alternatives if needed. This is called a “turnkey” service where you don’t do anything: just send them a BOM, let the fab deal with sourcing components.
As we were on a 10-boards run, I wasn’t going to deal with having to source a few 100nF capacitors to a fab in China; so I let them quote prices.
Prices quoted by the fab are a bit hit and miss. For instance they quoted US$3.30 for the LT3757 which is reasonable, but US$5.42 for the DA2032 which is absolutely outrageous (more on this later… but I can say in hindsight Coilcraft is the main culprit here). US$1.6 for the MOSFET was also acceptable. All in all, you’re definitely paying a premium but it is fair.
One thing I learned is that you mostly pay by the number of BOM lines. Having 2 different resistors instead of using twice the same weighs more in the final quotation. Initially PCBWay quoted a lot higher than their “$88 for 10” pricing probably because I was slightly exceeding what they deem reasonable for this promotional price. I discussed with them, changed the 100k pull-up on UVLO by 140k (also used for the frequency adjust pin) to save one line on the BOM and they proceeded to change the price.
Honestly, during the whole process they were amazing. I have only good things to say about PCBWay agents and no they are not paying me to say this!
I thought everything was all and well until they asked, in addition to the BOM and the Gerbers, a “centroid file”. I did not know what it was and again, I don’t think there is any industry standard for this. A quick googlefu and some exchanged emails later, they informed me it’s just a CSV file with the components’ center, and provided me with an example.
Under DipTrace it’s named “Pick and Place report”. Depending on your CAD tool of choice, your mileage may vary.
One last thing which is undoubtedly a lesson learned: don’t deviate from the norms: a few days after the process started I received an email from the fab’s agent asking for the orientation of D1. She pointed out the offender by sending me a picture of the silkscreen:
Seeing the first prototype work, it got me thinking. Do I really need to use this expensive LT3757 switching regulator? Is a flyback really better than a traditional boost converter if I use good parts for it? Should I test my theory that the dual diode is absolutely useless and a drag on performances?
Before taking this project further, I had to answer these questions.
So I decided to go ahead and launch a new set of prototypes, based on Texas Instruments’ LM3488. According to TI, the LM3488 is a “versatile low-side N-FET high-performance controller for switching regulators. This device is suitable for use in topologies requiring low-side FET, such as boost, flyback, or SEPIC.”; with a maximum output voltage of 500V.
It’s also a lot cheaper, and has simplified input/outputs with only 8 pins versus 11 for the LT3757. I thought to myself that maybe, just maybe, I was overengineering this power supply.
For this prototype I also included some jumpers that allows me to select different switching frequencies to see the effect it would have on efficiency.
The last prototype on the board is the same version as the first one, except it uses another transformer, the DA2033. It’s a bigger version of its little brother. Since I had space left on the PCB, I just wanted to experiment with this other transformer.
None of these designs gave me good results. To this date I don’t know why. Did I do something wrong? Are TI chips a bit more stringent on what they can accept? I do not know. However, it comforted me in my idea that the LT3757 was pretty darn nice in comparison.
After this brief interlude I worked on the revision B of the power supply.
A few improvements on this version:
When looking for a change of diode, I needed something FAST; with very little leakage. With a recovery time of 35ns and 1uA leakage @ 600V, this diode is a killer so I thought I’d give it a try.
On the picture above, don’t pay attention to the botched up job on the diode at the top, it was just an experiment I was running to see what was the impact of the dual diode versus the newly selected fast recovery diode. The answer to this question is: actually not so much difference. I still don’t understand why Linear’s engineers chose this more than average part in their reference design. There are so many good diodes out there, why this one? I doubt I’ll get an answer to this question.
Thanks to lots Shenzhen-based companies, producing PCBs is cheaper than it has ever been. Coupled with extremely aggressive pricing on shipping and you can basically get yourself professionally made PCBs delivered to your doorstep in about a week for less than US$30. Isn’t it a beautiful thing? Some people are still sad you can’t simply prototype on a breadboard with all these pesky SMD components. I say it’s a trade-off I am more than willing to accept! So the next logical step was to design a PCB.
This is the first design (Rev. A) I came up with:
Looking back, I see now straight away issues: the input voltage going through vias, ground and high voltage pins being way too close to each others; and this high frequency signal coming out of the transformer being dragged all around the board to the diode. As they say hindsight is 20/20.
First of all it’s amazing to see this thing work for the first time. In order to work with 0603 components and the tiny 0.5mm pitch regulator you absolutely need a microscope. It’s almost impossible to see solder bridges with the naked eye.
The other key ingredient to make this work is a stencil. Nowadays the PCB fab will ship you a stencil together with the board for a few more dollars. It’s just not practical to use solder paste with a syringe when you have pads so small. Everything will bridge and you will spend a ton of time fixing the board. Stencils are 100% worth their money.
The thing is, even with a stencil some 0.5mm pads will bridge up, because it’s applied by hand over the PCB stencils holes the amount of paste left will never be as good or as precise as a machine perfectly aligned with it. As with everything: practice makes perfect.
Reading 170V on the multimeter at the output is really reassuring and a good sign!
However, I could only see an efficiency of about 75%. It’s not awful considering the boost ratio we are dealing with, but it’s still miles away from what LTSpice had predicted. I found two culprits:
For the 2nd case, I simply unsoldered the 22 Ohm resistor part of the snubber network and efficiency immediately rose up to 80%+. For the diode, that will unfortunately need a revised prototype. It’s ok, I can improve the layout at the same time!
When it comes to design I trust the engineers from big semiconductors company. Why reinvent the wheel when it’s been done before; and by people that have a full time job in the field? So I took the 350V example from Linear, and adapted it to output the proper 170V Nixie voltage.
I even copied the weird double diode that they used; because I wanted first and foremost to have a working prototype before doing further changes. I just made sure the duty cycle would be OK with this new ratio and changed the feedback resistors but other than that: it is stock “typical application”.
The good thing about working with Linear products is that they are incredibly easy to simulate since they provide all the models on LTSpice. So simulation is the next logical step in the design phase:
I have a love/hate relationship with LTSpice. On one end it’s a powerful tool; on the other end its user experience from Windows 3.1 era could not be more frustrating to work with.
That being said, it is interesting to use because you can predict the sizing and power dissipation of components pretty accurately. For instance that’s how I detected that the 22 Ohm snubber can’t be a typical 1/10W 0603 resistor. It was absolutely beneficial to be able to simulate the design
Another interesting piece of information picked up from the simulation is predicting the voltage spikes at the drain of the transistor.
These spikes are cause by the inductance leakage in the transformer and will absolutely kill the MOSFET if it’s not properly avalanche rated. In LTSpice, this behavior can be simulated by using a less than 100% coupling coefficient. The DA2032 transformer specifies an inductance leakage of 0.150uH. That’s 0.150 / 10 = 0.015 or 1.5%. Rounded up to 2%, I used in the simulation a coupling coefficient of 1 – 0.02 = 0.98.
To be on the safe end, I decided in the end to go with a 100V rated MOSFET. Linear engineers went with 60V on their 350V application but I decided it doesn’t fit with the “build like a tank” design goal. Picking a MOSFET is not an easy task, there are thousands of reference out there. I used Mouser to quickly give an idea of what to look for:
Even with these restrictions, you are still left with over 200 results!
So now comes the dance of finding the right balance of low resistance, low gate charge, availability and cost. It is a painful process, but one that needs to be done and a good exercise in navigating the semiconductor industry’s jungle. In the end I found the one in the name of IRLR3110ZPbF. At 12mOhms Rds(on) and 34nC total gate charge, it is an extremely efficient switching MOSFET. In addition, it is rated “active and preferred” by Infineon meaning at the time of doing this prototype this part has no risk of being made obsolete in the near future.
The idea to build a good power high voltage DC/DC converter simply comes from my own need for it. Anything dealing with power management that comes from dodgy Chinese source is a potential for disaster. There has been some dramatic incident in the 3D printing community for instance. Add to this the high voltage element and I honestly would not leave for one second a nixie clock that is powered by these modules unattended.
Luckily for me, no drama happened: the power supply just crapped out without warning.
First and foremost let’s not reinvent the wheel. I like to think that there has to be a solution for this problem that is supported by a big semiconductor company. In the case of nixies though, not so much! If you study old designs, they’re all powered by mains and a big old transformer. A solution that is not hobby friendly.
That being said, there is at least one modern application that needs high voltage DC/DC transformation: capacitor chargers. They are typically the main source of power for photography using flash.
So after doing some more research, I found two key components that would fit the bill:
There are surprisingly not a whole lot of “stock” transformers out there, most of them are ordered custom made. However I dug up some interesting beauties from Coilcraft doing research:
I quickly found out that to generate high voltage, all semiconducators companies use a flyback transformer. No one uses a typical boost converter. The boost ratio is too great to have something safe and stable.
A transformer with 3/24V input used to generate 300V? Surely it can do 12 to 170V easily then right?
Coilcraft mentions LT3750 / LT3751 which are capacitor chargers, not switching regulators. They work on the same principle but have a few goodies like telling you when the capacitor is fully charged so that you can stop pumping high voltage.
Digging around though I came across the LT3757, a real switching regulator which includes a reference design for… a high voltage power supply!
The idea is all there, we are in business!
Would this work for a vacuum fluorescent display? I have a 128x64 Cherry display I got surplus, but it didn't come with the HV supply.
It's overkill for VFDs which need between 12 and 30 volts depending on model. Driving VFDs is a whole topic in itself.
You're not kidding… But I found a guide to the fundamentals: https://www.noritake-elec.com/technology/general-technical-information/vfd-operation
Nice job. I think a few extra dollars spent on a reliable power supply would be a good investment if you're powering a few hundred dollar's worth of new Dalibor Farny nixie tubes!
Tony
Jon
c.Invent
Brian Lough
Yann Guidon / YGDES
Thanks for spending the time documenting this. It was very interesting and informative.