Paralleling multiple ATX PSUs (and providing UPS)?

[IntuitiveNipple]

I have a large collection (15+) of surplus ATX PSUs and motherboards
of various ages currently doing nothing. I stripped them from their
original desktop/(mini)tower cases along with other components. I'm
now considering, as a project, making use of them in a highly
efficient and compact headless compute-cluster running Linux.

To that end I'm considering designing and building a custom power
supply solution. My idea is to create a rack with a single DC power
rail which each motherboard in the cluster will be connected to.

The ' power rail' will be powered from more than one ATX PSU with
their outputs connected in parallel to ensure sufficient current (with
suitable balancing circuitry between PSU outputs).

One benefit of this is that instead of having one PSU per motherboard,
each running nowhere near its maximum efficiency, I can calculate the
maximum current draw from all the motherboards and then use only the
number of PSUs sufficient to cover the maximum current draw. It would
also allow me to build multiple redundant 'power crates' that are hot-
swappable in the rack. If I use external connectors and circuitry to
link the PSUs to the rail (no interference with the PSUs internally),
any failed PSUs could be replaced in-situ quite easily.

Inspired by the Google server design that powers Gigabit GA-9IVDP
motherboards from a single 12V PSU output and a 12V SLA (sealed lead-
acid) battery per motherboard, I'm also thinking about using a bunch
of surplus ex-UPS 12V SLA batteries to keep the lights on over
temporary power glitches.

The pure 'Google' solution would require using something like the
picoPSU-120 ATX DC-DC device on each motherboard which takes a single
12V DC input (since the motherboards I have are all 20/24-pin ATX
3.3/5/12V types). This would make things simple since I'd only need to
parallel the +12V DC outputs from the ATX PSUs. Another benefit would
be ease of connection of renewable energy generators (wind/hydro). A
further benefit would be having the SLA batteries connected to the
'power rail' not individual motherboards.

The downside to this picoPSU design is having to spend quite a lot on
the devices (especially if I use all the motherboards!) as well as the
ATX PSUs not running at maximum efficiency (since their 5V/3V3 high-
current outputs would not be used) unless some of the 5V and/or 3.3V
outputs were connected serially and regulated to match the 12V
outputs.

I'd prefer to make use of all the outputs of the ATX PSUs *and*
provide the UPS functionality - talk about wanting your cake and to
eat it too. The only way I can imagine doing that right now is to hack
the PSUs such that the 12V DC batteries are connected to the PSU
common rectified DC output at a point where it is split into the
different output voltages. My guess is that it would be too
complicated especially as the bunch of PSUs are from various
manufacturers.
I suppose I could build custom regulators to provide 3V3/5V directly
from the 12V SLA batteries and have the batteries connected in
parallel to the 12V rail so they would charge when A/C was running the
PSUs.

My original aim was to reuse all this redundant equipment and in as
compact a space as possible - not waste masses of air-space by having
each mobo in its own case, and to make the power side as efficient as
possible (conversion efficiency required by the ATX12V standard is
around 70% so better to have 15 mobo's powered by 7 PSUs rather than
15 by 15). The mobo's will be mounted vertically to allow hot air to
escape easily with minimal blowing.

I'm wondering if anyone else has been crazy enough to attempt
something like this with power supplies? I've seen a lot of articles
from my research that show that other people have successfully
implemented most of what I'm aiming for as discrete power projects but
no one has attempted to 'join the dots' that I can find.

 

[frischmoutt]

I have a large collection (15+) of surplus ATX PSUs and motherboards
of various ages currently doing nothing. I stripped them from their
original desktop/(mini)tower cases along with other components. I'm
now considering, as a project, making use of them in a highly
efficient and compact headless compute-cluster running Linux.

This post smells the troll. However, i'll reply

Forgetaboutit. They're not designed to be connected in //.
Their own regulations are not tuned on the same value and their output
resistances (static as well as dynamic ones) are not the same.
This means that the overall current won't be distributed in an equitable
way. One being overloaded whilst the other will deliver a poor current.
The only way to ensure the current sharing would be to put series resistors
in every output. Silly solution.
On top of that , they're not using the same converter design, not the same
frequencies, bla bla. Interferences will most probably cause malfunctions.

IMHO, practically, the best way is to keep them completely separated and to
build a circuit (multi-contact relay for example) allowing to start all of
them at the same moment.
Again, technically this is probably not a good solution because these PSU
use a bridge rectifier in their inputs followed by a big filtering capacitor
connected to it. Turning them On at the same instant will draw a hudge
current peak. Unless your circuit braker is able to withstannd it or you
aren't starting the computer during the night, you can make sure that you'll
be in the dark !

Good luck

 

[Paul]

I have a large collection (15+) of surplus ATX PSUs and motherboards
of various ages currently doing nothing. I stripped them from their
original desktop/(mini)tower cases along with other components. I'm
now considering, as a project, making use of them in a highly
efficient and compact headless compute-cluster running Linux.

To that end I'm considering designing and building a custom power
supply solution. My idea is to create a rack with a single DC power
rail which each motherboard in the cluster will be connected to.

The ' power rail' will be powered from more than one ATX PSU with
their outputs connected in parallel to ensure sufficient current (with
suitable balancing circuitry between PSU outputs).

One benefit of this is that instead of having one PSU per motherboard,
each running nowhere near its maximum efficiency, I can calculate the
maximum current draw from all the motherboards and then use only the
number of PSUs sufficient to cover the maximum current draw. It would
also allow me to build multiple redundant 'power crates' that are hot-
swappable in the rack. If I use external connectors and circuitry to
link the PSUs to the rail (no interference with the PSUs internally),
any failed PSUs could be replaced in-situ quite easily.

Inspired by the Google server design that powers Gigabit GA-9IVDP
motherboards from a single 12V PSU output and a 12V SLA (sealed lead-
acid) battery per motherboard, I'm also thinking about using a bunch
of surplus ex-UPS 12V SLA batteries to keep the lights on over
temporary power glitches.

The pure 'Google' solution would require using something like the
picoPSU-120 ATX DC-DC device on each motherboard which takes a single
12V DC input (since the motherboards I have are all 20/24-pin ATX
3.3/5/12V types). This would make things simple since I'd only need to
parallel the +12V DC outputs from the ATX PSUs. Another benefit would
be ease of connection of renewable energy generators (wind/hydro). A
further benefit would be having the SLA batteries connected to the
'power rail' not individual motherboards.

The downside to this picoPSU design is having to spend quite a lot on
the devices (especially if I use all the motherboards!) as well as the
ATX PSUs not running at maximum efficiency (since their 5V/3V3 high-
current outputs would not be used) unless some of the 5V and/or 3.3V
outputs were connected serially and regulated to match the 12V
outputs.

I'd prefer to make use of all the outputs of the ATX PSUs *and*
provide the UPS functionality - talk about wanting your cake and to
eat it too. The only way I can imagine doing that right now is to hack
the PSUs such that the 12V DC batteries are connected to the PSU
common rectified DC output at a point where it is split into the
different output voltages. My guess is that it would be too
complicated especially as the bunch of PSUs are from various
manufacturers.
I suppose I could build custom regulators to provide 3V3/5V directly
from the 12V SLA batteries and have the batteries connected in
parallel to the 12V rail so they would charge when A/C was running the
PSUs.

My original aim was to reuse all this redundant equipment and in as
compact a space as possible - not waste masses of air-space by having
each mobo in its own case, and to make the power side as efficient as
possible (conversion efficiency required by the ATX12V standard is
around 70% so better to have 15 mobo's powered by 7 PSUs rather than
15 by 15). The mobo's will be mounted vertically to allow hot air to
escape easily with minimal blowing.

I'm wondering if anyone else has been crazy enough to attempt
something like this with power supplies? I've seen a lot of articles
from my research that show that other people have successfully
implemented most of what I'm aiming for as discrete power projects but
no one has attempted to 'join the dots' that I can find.

It's not possible to answer a question such as this, without writing a book.
I'm sure a power engineer would be willing to entertain your question -- for
about $200/hour in consulting fees.

I think you'll find, for any reasonably small project (just the one rack),
using existing retail solutions is more than adequate. Even your home
insurance agent will be proud of you, for using UL approved solutions,
and might even pay out on a fire claim.

With regard to the notion of making a large 12V, consider that there
is a difference between a "distribution" voltage and a "regulated load"
voltage. Attempting to cross those two functions, such as Google has
done, would require a good deal of care. Mixing the two functions,
produces a Less regulated voltage. There is a danger, if you do it
that way, that the hard drive will spin down and spin up again, over
and over again. Hard drives are sensitive to the voltages fed to them,
more so than a Vcore regulator on a motherboard would be. The motherboard
nay be more tolerant of a sloppy 12V feed.

Paul