A relevant part that changed my view of charge pumps is the LTC7820 [0]. This is an inductorless charge pump that can be used as a an unregulated voltage doubler or halver... at 500+ W and 98%+ efficiency. I used to think of charge pumps as designed for generating bias voltages where the actual power is quite small... but this shows that they scale quite well. (There's also the LTC7821 that combines the unreglated inductorless halver of the '7820 with a regulated, nominally-2:1 buck to give a regulated 48V -> 12V converter with some impressive efficiency numbers.)
These circuits take a lot of parts to do a job that you can do with modern high frequency stuff with a lot lower cost and parts count.
The normal point of a capacitive doubler is either to give you a voltage you need without a lot of extra parts count (often negative) or to generate a very high voltage.
The app notes and data sheets of related parts suggest that the target application is large conversion ratios, where the duty cycle in an inductive converter is close to 0 (or 1). That forces tradeoffs, like a lower switching frequency (lower efficiency), a larger inductor (more weight and/or cost) or very short T_on for one of the FETs (lower efficiency because transition times become important). So you can use the charge pump as the first stage of a hybrid converter to get a higher system efficiency.
My recommendation to anyone who finds themself stuck in this corner of the design space is to consider a tapped inductor converter.
I did a tapped inductor boost last year to take 3V input to 80V output (at not-much output current, I forget exactly what it was but it was mostly a bias voltage; also, the actual output voltage was DAC-set and could be quite low, so the loop dynamics were unpleasant). It was definitely annoying to wrap my head around, and very annoying to select the inductor (Würth has a nice OTS series, at the usual Würth prices; HVM would likely want a custom or semicustom design) but it just plain worked the first try and continued working through the usual stress tests and also the unusual stress tests of the Very Expensive Load™ getting itself Very Expensively Killed™ (for non-power-supply reasons). I was really happy with that converter, that kind of step-up ratio isn't easy and it just worked.
My refs were written into the schematic so I dug it up and might as well post the whole thing (SEE WARNINGS BELOW): https://i.ibb.co/zVXZWxHg/tapped-boost.png (forgive me, I have no idea on the state of image hosting in 2025 so I hope that link holds up)
I misremembered what design I actually ended up building. It was nominal 5V to 75V but was tested at a wider range (including the mentioned 3V to 80V). The variable voltage was implemented later, in a linear stage, which did in fact have a number of issues.
ADI app note AN-1126 was very helpful in this design. It pushes for a different topology, yet compares with this one and others in some detail. Its arguments against tapped inductor designs are threefold: (1) sometimes EMI issues on the switching node (fair, and a real concern, but not interesting for what I needed to do), (2) the inductor is annoying to source (very true), and (3) the main technical objection is that demands on the output rectifier are high and might require you to use something crappier than a Schottky. That last one is true in general but for this design in this decade, I was comfortably within medium voltage Schottky territory, and so their main objection was a complete nonissue. That looked good to me so I went ahead and built it, and was not disappointed.
The inductor was 744889030330 (say that three times fast) from the WE-MTCI family which worked a dream. I don't remember why this exact switcher chip was chosen but I do remember having a lot of candidates and so the choice was kind of arbitrary. The control scheme type it uses is important, though. I don't think the zener did anything but being paranoid I wanted to have the footprint there for the prototype build rather than not have it.
THE PREVIOUSLY PROMISED WARNINGS WHICH ARE ACTUALLY REALLY IMPORTANT SO I AM USING CAPS:
1. 80V can kill you. Really, it can. Use caution!
2. This is the design I sent off for fab (and perhaps not even the final version, if the folder notes are any guide). It is not the redlined, working version. I believe this stage was OK. Maybe it needed a bit more output capacitance? Certainly I know the following stage, not shown here, had severe issues (most of them stupid and obvious once noticed). So treat this as good inspiration, not ground truth.
3. Did I mention that 80V can kill you? USE CAUTION.
Thanks for the refs and discussion--especially that app note.
I will note that we're all talking about 10+ year old chips and technology--mostly prior to GaN.
Apparently, I'm going to have to dig through a bunch of stuff to see what is current. While the topologies don't change, the tradeoffs between them do as technology changes.
Edit: Stare at the LT7890/1/2/3 series for GaN stuff by way of comparison
GaN is not appropriate for most use cases. It matters when high frequency matters, which is usually when magnetics size matters. I have also seen it used very well for ultra-low input voltages (basically energy harvesting).
It is also uncompromising and brutally difficult to get working well. GaN-FETs love to commit suicide in new and entertaining ways. And LTC7890 is not something I would want to implement in any design (though of course I'd suck it up and do it if it was the right choice).
99% of the market is traditional boring stuff because 99% of the market is well served by traditional boring stuff.
ZVS (not ZCS) topologies that use planned stray inductance and WBG transistors and class 1 ceramics can operate at 10s of MHz with less than 1% of loss per voltage halving.
They have insane power density and lack of wound inductors means there's nothing that causes problems as you push into the kilovolts.
But really they're just more efficient (integer ratios, where `n-1` not prime) than classical Buck/boost topologies.
GaN stuff can be 99%+ efficiency. The frequencies are multiple MHz which shrinks the inductors significantly--sometimes allowing PCB based coils (See Anker 120W teardowns).
I'm a huge fan of GaN (see the current front page discussion of LiDAR), but 99% at 500 W at 12 V is hard. 3 milliohm in the inductor will eat up 100% of your budget, without accounting for any switching losses. Like most Linear parts, the LTC7820 is impressive, fits a niche, and is too expensive for most applications -- but when you need it, it's good to know it's there.
Yeah 120W is table stakes.
Also Anker wall adapters are insulated that's worse for efficiency.
Integer ratios can be done with GaN and class 1 ceramics and stray inductance to charge the GaN drain, so they only experience ZVS (they don't need ZCS).
Use a resonant gate drive.
Efficiencies around 99.8% are feasible for a doubler/halver based on the "3L FC buck" topology, in the 10~30 MHz range.
I’m with you, I’m not sure the volume or cost would be less once you factor in capacitors that are high enough quality for the application.
The datasheet mentions low profile a lot. That does make sense as one can make a flat, high quality, capacitor. Making a flat high quality inductor is harder and probably more expensive and likely consumes more volume overall. I can imagine some applications where being flat is important, like the back of a panel.
Is this really even a little cursed? It's a perfectly nice device, and in fact has been manufactured as a chip before. Enpirion has made this under the name of EC2650QI 6A Voltage Divider:
Enpirion's first products - integrated-inductor DC/DC converters - were limited to fairly low input voltages. This allowed them to run very fast with the technology of the time; they needed to run fast, because the integrated inductor was not very large. But this was severely limiting Enpirion's market: if you wanted to make something 5V-powered, they were great. But a PC motherboard application with 12V input?
This device was the answer: convert 12V to 6V with this switched capacitor halver, and then use other Enpirion parts. I don't think it was super successful, though, because at this point you were no longer winning on board complexity by using integrated inductors.
I think I (a long time software nerd) am finally getting over the hump with analog stuff. The article says the circuit is complicated and hard to understand, yet I got it instantly. Feels sort of like learning a musical instrument and realizing that one of the early pieces you struggled with is easy now.
FWIW: the didactic trick of imagining the floating capacitor "carrying" charge from one "place" to another was really good. That's not the way most treatments talk about charge pumps, and I think it's a lot cleaner.
It's not complicated. It's just very basic capacitor behavior. If there is a tricky part here then it is in the bit that is glossed over: the switches. But congrats on getting it! Analog is fun, you can get incredibly complex behavior out of a handful of components.
> It's not complicated. It's just very basic capacitor behavior.
All circuits are just very basic circuit element behavior. In fact a charge pump is a decidedly counter-intuitive thing, up there with things like "why a long-tailed pair makes a differential amplifier" or "how tf does a buck/boost regulator work?".
This is like looking at a 30-line function implementing a FFT and announcing "it's just very basic C code".
I don't know about that. I was building inverters from scratch on intuition when I was 17 or so and I started with electronics properly when I was 11 reading library books, taking stuff apart and putting it (or something else) back together again.
Compared to an FFT, which would require a lot of math insight it seems to me a charge pump is trivial because charging and discharging is what capacitors do, it is the most basic thing you can do with a cap. In programming it would not be the equivalent of an FFT, but more like the equivalent of putting a value in a variable and expecting that value to still be there a while later.
Now, I came into software from hardware, to me software is having an infinite parts box so what's trivial to someone that started in hardware is probably entirely different from the view a software person has of the hardware world. Which is one reason why analog is such a barrier to programmers: they tend to look for the place where the state of the system is stored, rather than that the whole circuit is a continuous function of its input and what happened before. Getting analog circuitry to behave predictably can be quite tricky because of that and just hanging a probe of it will influence the circuit.
Resistors are easy, for the most part grade school math. Capacitors harder but not a lot harder, high school math, some basic integration required. Coils are where it gets more interesting and difficult math wise. And of course parasitic components are everywhere and can make your life harder, especially at higher frequencies or, in the case of single pulses shorter rise and fall times. A piece of wire is a resistor, capacitor and a coil for free all at once. Stripline takes advantage of that fact but without tooling it is next to impossible to work out the math and get predictable results (again, for me, I know people that can do this intuitively).
Multi-pole filters, high gain amplifiers, clean oscillators, those rise to the level of that FFT (for me).
The long tailed pair I agree with, that's a puzzler.
If you don't have one yet (and I assume you do, but just in case): get a scope. They're pretty cheap nowadays and a good scope will do wonders for such insights. I only got one when I was an adult and finally could afford it and it was an absolute game changer. If I had had one earlier I would have gotten a lot more mileage out of my time, especially when troubleshooting broken hardware. You don't need a fancy one, 100 MHz dual trace is more than good enough but that's table stakes now.
One big insight for me is that hardware 'is' and software 'does'. You don't tell a piece of hardware what to do, you design it to 'be' something and a computer simply (ok, maybe not so simply) needs to be told what to 'do'. This took me a long time to really grok after making the jump from hardware to software.
I was trying to figure out where the cursed part comes into it; I assume it's the switching, that typical circuit analysis just doesn't involve discrete states that are swapped between?
Yep. The state where the capacitor is fully floating relies on non-obvious details. Recall that a (real) transistor has a rated maximum voltage that it can block. But when the (ideal) transistors are opened, there's no defined voltage relationship between (either side of) the floating capacitor and the rest of the circuit; so if analyzed at this level, its not clear if the transistors are within spec or not. You need to at least think about the parasitics to convince yourself that this is sane... or build it with real parts and see that it works.
Indeed. That's why the MAX232 is such an impressive little chip, it was the first time I found an on-chip charge pump in the wild. One of the more interesting uses for charge pumps is loss-free balancing of battery cells. I really like it because contrary to resistive balancing it works with the charge already in the cells by redistributing it.
It seems as if you could get a range of voltages at the point between C1 and C2 with a closed negative feedback loop that controls the duty cycle of the flying capacitor.
If you charge the bottom cap more and the top one less, you can jack the voltage toward the power rail.
A buck as well as boost-buck converter could be produced without inductors.
The article references the LTC3265 IC, whose datasheet says "The LDO output voltages can be adjusted using external resistor dividers" (connected to the ADJ pins).
If I remember correctly, some kind of highly efficient (and low cost) voltage halver (not sure if this design or a different one) is AFAIK used with the PPS ("programmable power supply") protocol to let phones charge efficiently:
2x the battery charge voltage is requested from the power supply, e.g. 8.6V if the phone is trying to apply 4.3V to the battery. This way, the phone doesn't need to run any complicated and heat-generating voltage regulation, just the halver, while still being able to request more than the standard 5V over the cable (allowing it to draw more than 15W of power over a standard cable).
You can e.g. stack two half bridges, connect their switching nodes with a flying capacitor, and declare the connection of the stack the new switching node. If you use a series induction from the switching node to the load, that's a "3L FC buck". The inductor is not needed for halving.
Fascinating design I haven't tried, I have made inductor based designs but a pure capacitor design combined with a high speed mos might make for a fun micro psu design.
These types of switching circuits are very common inside ASIC where the high speed isn’t an issue, you don’t need to move all that much charge, and one can’t easily (if at all) support inductors.
I can't recall if it was Falstad or which other simulator code I read, but it had "connect this one terminal to ground via multi-Gigaohm resistor for stability" sprinkled throughout the code for capacitors and similar components.
A relevant part that changed my view of charge pumps is the LTC7820 [0]. This is an inductorless charge pump that can be used as a an unregulated voltage doubler or halver... at 500+ W and 98%+ efficiency. I used to think of charge pumps as designed for generating bias voltages where the actual power is quite small... but this shows that they scale quite well. (There's also the LTC7821 that combines the unreglated inductorless halver of the '7820 with a regulated, nominally-2:1 buck to give a regulated 48V -> 12V converter with some impressive efficiency numbers.)
[0] https://www.analog.com/media/en/technical-documentation/data...
But, why use those parts?
These circuits take a lot of parts to do a job that you can do with modern high frequency stuff with a lot lower cost and parts count.
The normal point of a capacitive doubler is either to give you a voltage you need without a lot of extra parts count (often negative) or to generate a very high voltage.
The app notes and data sheets of related parts suggest that the target application is large conversion ratios, where the duty cycle in an inductive converter is close to 0 (or 1). That forces tradeoffs, like a lower switching frequency (lower efficiency), a larger inductor (more weight and/or cost) or very short T_on for one of the FETs (lower efficiency because transition times become important). So you can use the charge pump as the first stage of a hybrid converter to get a higher system efficiency.
My recommendation to anyone who finds themself stuck in this corner of the design space is to consider a tapped inductor converter.
I did a tapped inductor boost last year to take 3V input to 80V output (at not-much output current, I forget exactly what it was but it was mostly a bias voltage; also, the actual output voltage was DAC-set and could be quite low, so the loop dynamics were unpleasant). It was definitely annoying to wrap my head around, and very annoying to select the inductor (Würth has a nice OTS series, at the usual Würth prices; HVM would likely want a custom or semicustom design) but it just plain worked the first try and continued working through the usual stress tests and also the unusual stress tests of the Very Expensive Load™ getting itself Very Expensively Killed™ (for non-power-supply reasons). I was really happy with that converter, that kind of step-up ratio isn't easy and it just worked.
Links or refs? That sounds very interesting.
A basic search coughs up a bunch of papers from academic paper mills, and I don't see obvious links to an OTS series from Wurth Elektronik.
My refs were written into the schematic so I dug it up and might as well post the whole thing (SEE WARNINGS BELOW): https://i.ibb.co/zVXZWxHg/tapped-boost.png (forgive me, I have no idea on the state of image hosting in 2025 so I hope that link holds up)
I misremembered what design I actually ended up building. It was nominal 5V to 75V but was tested at a wider range (including the mentioned 3V to 80V). The variable voltage was implemented later, in a linear stage, which did in fact have a number of issues.
ADI app note AN-1126 was very helpful in this design. It pushes for a different topology, yet compares with this one and others in some detail. Its arguments against tapped inductor designs are threefold: (1) sometimes EMI issues on the switching node (fair, and a real concern, but not interesting for what I needed to do), (2) the inductor is annoying to source (very true), and (3) the main technical objection is that demands on the output rectifier are high and might require you to use something crappier than a Schottky. That last one is true in general but for this design in this decade, I was comfortably within medium voltage Schottky territory, and so their main objection was a complete nonissue. That looked good to me so I went ahead and built it, and was not disappointed.
The inductor was 744889030330 (say that three times fast) from the WE-MTCI family which worked a dream. I don't remember why this exact switcher chip was chosen but I do remember having a lot of candidates and so the choice was kind of arbitrary. The control scheme type it uses is important, though. I don't think the zener did anything but being paranoid I wanted to have the footprint there for the prototype build rather than not have it.
THE PREVIOUSLY PROMISED WARNINGS WHICH ARE ACTUALLY REALLY IMPORTANT SO I AM USING CAPS:
1. 80V can kill you. Really, it can. Use caution!
2. This is the design I sent off for fab (and perhaps not even the final version, if the folder notes are any guide). It is not the redlined, working version. I believe this stage was OK. Maybe it needed a bit more output capacitance? Certainly I know the following stage, not shown here, had severe issues (most of them stupid and obvious once noticed). So treat this as good inspiration, not ground truth.
3. Did I mention that 80V can kill you? USE CAUTION.
Thanks for the refs and discussion--especially that app note.
I will note that we're all talking about 10+ year old chips and technology--mostly prior to GaN.
Apparently, I'm going to have to dig through a bunch of stuff to see what is current. While the topologies don't change, the tradeoffs between them do as technology changes.
Edit: Stare at the LT7890/1/2/3 series for GaN stuff by way of comparison
GaN is not appropriate for most use cases. It matters when high frequency matters, which is usually when magnetics size matters. I have also seen it used very well for ultra-low input voltages (basically energy harvesting).
It is also uncompromising and brutally difficult to get working well. GaN-FETs love to commit suicide in new and entertaining ways. And LTC7890 is not something I would want to implement in any design (though of course I'd suck it up and do it if it was the right choice).
99% of the market is traditional boring stuff because 99% of the market is well served by traditional boring stuff.
ZVS (not ZCS) topologies that use planned stray inductance and WBG transistors and class 1 ceramics can operate at 10s of MHz with less than 1% of loss per voltage halving.
They have insane power density and lack of wound inductors means there's nothing that causes problems as you push into the kilovolts.
But really they're just more efficient (integer ratios, where `n-1` not prime) than classical Buck/boost topologies.
Inductors are large, expensive, hot, and unreliable.
Not on modern high-frequency switchers.
GaN stuff can be 99%+ efficiency. The frequencies are multiple MHz which shrinks the inductors significantly--sometimes allowing PCB based coils (See Anker 120W teardowns).
I'm a huge fan of GaN (see the current front page discussion of LiDAR), but 99% at 500 W at 12 V is hard. 3 milliohm in the inductor will eat up 100% of your budget, without accounting for any switching losses. Like most Linear parts, the LTC7820 is impressive, fits a niche, and is too expensive for most applications -- but when you need it, it's good to know it's there.
Yeah 120W is table stakes. Also Anker wall adapters are insulated that's worse for efficiency.
Integer ratios can be done with GaN and class 1 ceramics and stray inductance to charge the GaN drain, so they only experience ZVS (they don't need ZCS). Use a resonant gate drive. Efficiencies around 99.8% are feasible for a doubler/halver based on the "3L FC buck" topology, in the 10~30 MHz range.
I’m with you, I’m not sure the volume or cost would be less once you factor in capacitors that are high enough quality for the application.
The datasheet mentions low profile a lot. That does make sense as one can make a flat, high quality, capacitor. Making a flat high quality inductor is harder and probably more expensive and likely consumes more volume overall. I can imagine some applications where being flat is important, like the back of a panel.
> you can do with modern high frequency stuff
EMC
The '7660 is good for low-power and is my go-to DIP-8 part when I need a half or double voltage supply on a breadboard.
Is this really even a little cursed? It's a perfectly nice device, and in fact has been manufactured as a chip before. Enpirion has made this under the name of EC2650QI 6A Voltage Divider:
https://cdrdv2-public.intel.com/632833/ec2650qi-datasheet.pd...
Enpirion's first products - integrated-inductor DC/DC converters - were limited to fairly low input voltages. This allowed them to run very fast with the technology of the time; they needed to run fast, because the integrated inductor was not very large. But this was severely limiting Enpirion's market: if you wanted to make something 5V-powered, they were great. But a PC motherboard application with 12V input?
This device was the answer: convert 12V to 6V with this switched capacitor halver, and then use other Enpirion parts. I don't think it was super successful, though, because at this point you were no longer winning on board complexity by using integrated inductors.
I think I (a long time software nerd) am finally getting over the hump with analog stuff. The article says the circuit is complicated and hard to understand, yet I got it instantly. Feels sort of like learning a musical instrument and realizing that one of the early pieces you struggled with is easy now.
FWIW: the didactic trick of imagining the floating capacitor "carrying" charge from one "place" to another was really good. That's not the way most treatments talk about charge pumps, and I think it's a lot cleaner.
It's not complicated. It's just very basic capacitor behavior. If there is a tricky part here then it is in the bit that is glossed over: the switches. But congrats on getting it! Analog is fun, you can get incredibly complex behavior out of a handful of components.
> It's not complicated. It's just very basic capacitor behavior.
All circuits are just very basic circuit element behavior. In fact a charge pump is a decidedly counter-intuitive thing, up there with things like "why a long-tailed pair makes a differential amplifier" or "how tf does a buck/boost regulator work?".
This is like looking at a 30-line function implementing a FFT and announcing "it's just very basic C code".
I don't know about that. I was building inverters from scratch on intuition when I was 17 or so and I started with electronics properly when I was 11 reading library books, taking stuff apart and putting it (or something else) back together again.
Compared to an FFT, which would require a lot of math insight it seems to me a charge pump is trivial because charging and discharging is what capacitors do, it is the most basic thing you can do with a cap. In programming it would not be the equivalent of an FFT, but more like the equivalent of putting a value in a variable and expecting that value to still be there a while later.
Now, I came into software from hardware, to me software is having an infinite parts box so what's trivial to someone that started in hardware is probably entirely different from the view a software person has of the hardware world. Which is one reason why analog is such a barrier to programmers: they tend to look for the place where the state of the system is stored, rather than that the whole circuit is a continuous function of its input and what happened before. Getting analog circuitry to behave predictably can be quite tricky because of that and just hanging a probe of it will influence the circuit.
Resistors are easy, for the most part grade school math. Capacitors harder but not a lot harder, high school math, some basic integration required. Coils are where it gets more interesting and difficult math wise. And of course parasitic components are everywhere and can make your life harder, especially at higher frequencies or, in the case of single pulses shorter rise and fall times. A piece of wire is a resistor, capacitor and a coil for free all at once. Stripline takes advantage of that fact but without tooling it is next to impossible to work out the math and get predictable results (again, for me, I know people that can do this intuitively).
Multi-pole filters, high gain amplifiers, clean oscillators, those rise to the level of that FFT (for me).
The long tailed pair I agree with, that's a puzzler.
If you don't have one yet (and I assume you do, but just in case): get a scope. They're pretty cheap nowadays and a good scope will do wonders for such insights. I only got one when I was an adult and finally could afford it and it was an absolute game changer. If I had had one earlier I would have gotten a lot more mileage out of my time, especially when troubleshooting broken hardware. You don't need a fancy one, 100 MHz dual trace is more than good enough but that's table stakes now.
One big insight for me is that hardware 'is' and software 'does'. You don't tell a piece of hardware what to do, you design it to 'be' something and a computer simply (ok, maybe not so simply) needs to be told what to 'do'. This took me a long time to really grok after making the jump from hardware to software.
I was trying to figure out where the cursed part comes into it; I assume it's the switching, that typical circuit analysis just doesn't involve discrete states that are swapped between?
Yep. The state where the capacitor is fully floating relies on non-obvious details. Recall that a (real) transistor has a rated maximum voltage that it can block. But when the (ideal) transistors are opened, there's no defined voltage relationship between (either side of) the floating capacitor and the rest of the circuit; so if analyzed at this level, its not clear if the transistors are within spec or not. You need to at least think about the parasitics to convince yourself that this is sane... or build it with real parts and see that it works.
Indeed. That's why the MAX232 is such an impressive little chip, it was the first time I found an on-chip charge pump in the wild. One of the more interesting uses for charge pumps is loss-free balancing of battery cells. I really like it because contrary to resistive balancing it works with the charge already in the cells by redistributing it.
FETs are capacitors that can be internally shorted out.
A good model for charge pump and duct switching converter understanding.
It seems as if you could get a range of voltages at the point between C1 and C2 with a closed negative feedback loop that controls the duty cycle of the flying capacitor.
If you charge the bottom cap more and the top one less, you can jack the voltage toward the power rail.
A buck as well as boost-buck converter could be produced without inductors.
Indeeed, I found an article about exactly this: https://www.allaboutcircuits.com/technical-articles/boosting...
The article references the LTC3265 IC, whose datasheet says "The LDO output voltages can be adjusted using external resistor dividers" (connected to the ADJ pins).
If I remember correctly, some kind of highly efficient (and low cost) voltage halver (not sure if this design or a different one) is AFAIK used with the PPS ("programmable power supply") protocol to let phones charge efficiently:
2x the battery charge voltage is requested from the power supply, e.g. 8.6V if the phone is trying to apply 4.3V to the battery. This way, the phone doesn't need to run any complicated and heat-generating voltage regulation, just the halver, while still being able to request more than the standard 5V over the cable (allowing it to draw more than 15W of power over a standard cable).
Yeah PPS uses capacitive halvers.
You can e.g. stack two half bridges, connect their switching nodes with a flying capacitor, and declare the connection of the stack the new switching node. If you use a series induction from the switching node to the load, that's a "3L FC buck". The inductor is not needed for halving.
Fascinating design I haven't tried, I have made inductor based designs but a pure capacitor design combined with a high speed mos might make for a fun micro psu design.
These types of switching circuits are very common inside ASIC where the high speed isn’t an issue, you don’t need to move all that much charge, and one can’t easily (if at all) support inductors.
A simulator will certainly not like the floating center terminal.
I just threw it into Micro-Cap and it surprisingly didn't throw any errors with the floating node. https://imgbox.com/riKCyWI5
I can't recall if it was Falstad or which other simulator code I read, but it had "connect this one terminal to ground via multi-Gigaohm resistor for stability" sprinkled throughout the code for capacitors and similar components.
SPICE can do that with 'gshunt' option.
Could be also done with 2 switches and some PWM closed loop control. Which you probably want to do anyway as the halving won't hold under load.