General plan view of ATX-style, convection cooled, PC Power Supply.
David J Greaves. University of Cambridge, March 2020.
I wanted the world's quietest power supply for my PC. The best way to achieve this is convection air cooling with no fan. Compared with previous decades, many modern PCs do not use a great deal of power, so cooling the power supply should not really be a problem. Many PC cases, however, use the power supply fan as the main case fan and use air vents with low cross-sectional area. This causes greater wind speed and hence noise. So the most important step is to use a very large case with very large vents, or, as I do here, take the power supply out of the PC case and get greater access to cool air that way.
A power supply gives off three types of noise. I discuss each of them in this research note. 1/ There is airborne sound and physical vibration arising from air movement and other moving parts. 2/ There is electromagnetic emission (EMI) radio waves that can intefere with wireless systems such as broadcast radio, DECT, WiFi and Bluetooth. 3/ There is earth current that can interfere with audio electronics such as microphone and amplifier systems.
The key to avoiding airborne noise is to reduce heat density and air speed. The primary approach is to increase heat sink area so that forced air is not needed and hence no fan or other moving parts are needed. In this project, I created an external PSU, using four off-the-shelf SMPSUs, bolted to a slotted steel plate, and suspended behind the desk next to the PC enclosure. By reducing the heat generation inside the PC enclosure, PC cooling is easier and almost fanless cooling is possible using a very large, heat-piped, passive CPU heatsink. I used a Noctua NH-U9B SE2, Premium CPU Kühler:
This deadly-quiet ATX PSU was constructed using four, floating SMPSU modules to provide the +5, +12, +3.3 and +5 standby rails. The -12 rail was generated from a linear power supply and a 7912 linear regualtor. Current draw on this rail is typically very much smaller than the 500mA specification. This transformer's other purpose is to generate a quick-die mains good signal. A low-value smoothing capacitor is used in a second output path from this transformer. This output decays much faster than the main rails when the mains power is removed and hence enables the ATX `power good' signal to be deasserted well before the supply fails.
Modern CPUs are less power hungry than a few years ago. There is also a shift to taking much more of the power from the 12V rail. For the same power, the 12V rail requires thinner wires and wastes less power in the rectifier diodes which have the same drop regardless of rail potential being generated, but which dissipate less power at lower currents (greater-than linear saving owing to drop being less at less current).
The current needs of a motherboard can be predicted by the Outervision web site: LINK. An Intel I7 with two 1TB SSD drives and 16 GByte of DRAM is reckoned to need 8A each on 3V3 and 5V and 4.4A on the 12V rail. This comes to 170 watts. The component PSUs I used were made by Meanwell and were all rated at about 16A so will be well within specification.
As well as feeding the PC, I added a second 5V standby output for USB hubs. Generally I would prefer to power off the USB hubs when the main power goes off, but I've found most modern motherboards no longer give the option to turn off USB sockets when main power is off, and if the hubs are powered by the main 5V rail, an unwanted path is created. The unwanted path is that the 5V standby rail enters the PC and comes out of the USB sockets; it then passes to a powered hub that I have connected to the 5V rail of the PC and current flows the wrong way down this path when main power is off.
I used a slotted steel backplate to mount all components on. This formed the bottom of my enclosure, with convection air current passing through the slots and around both ends of the backplate.
Each component PSU has two sets of heat-generating components that are thermally bonded to the metal backplate. These are the power switching MOSFETs in the primary circuit and there are schottky rectifier diodes in the secondary circuit. Care should be taken with the orientation of mounting the component PSUs. The diodes generate more heat than the MOSFETs. Therefore, each PSU should be bolted to the main backplate such that the rectifiers have the shortest thermal path, but also taking into account the airflow design so that air can adequately convect through the vents on the component PSU. The rectifiers and MOSTFETs are both likely to be in TO-220 cases, but can be readily distinguished by looking at the general layout of the component PSU: the rectifiers will be closer to the DC output terminals and the MOSFETs will be closer to the AC input terminals.
The block diagram and heavy current wiring is shown in the above figure. An IEC inlet socket with integral 20mm fuse provides primary protection for the linear supply. If a larger supply is built to this design, a second 20mm fuse should be provided for the transformer, as shown. I used an IEC inlet connector that also had an integral mains filter. The filter is not really needed, but the part was to hand. I put a 4A mains fuse in it. A single-pole mains on/off switch is inserted in the live. The standby supply is always on while the mains is ON. The other four loads are switched using a 250V 25A solid-state relay. This is overspecified, but was at hand.
The ATX specification requires that the 3V3 rail is never greater than the 12 or 5 volt rail in potential. This is (sort of) guaranteed by the additional 1N540X power diodes. These are rated at 3A continuous, but are specified to handle 200 amps in surge situations, which is what is required here.
Power supply sequencing and monitoring is provided by an EST7502C IC (PDF). This was removed from a failed ATX power supply which also provided the molex connectors for the motherboard and SATA SSDs.
This controller IC contains an oscillator, voltage comparator and PWM modulator to generates signals for a complementary pair of switching MOSFETs. These functions are not needed in my application. I do use the remaining functions, which are to receive the ON signal from the PC, provide soft start and rail monitoring and generate the POWER GOOD signal to the PC. The principle difference in configuration is that no feedback voltage monitoring was needed (commonly taken from a light grey sense wire at the main modex connector) and that one of the two open-collector PWM outputs was simply low-pass filtered and used to turn on the solid-state relay. I could have paralleled the two outputs I think.
The umbilical cable between power supply and PC raises two concerns: voltage drop and radiation.
The voltage drop can be mitigated by using one half to one third of a standard umbilical and thickly extending every wire to the required length. By using only a fraction of an existing umbilcal and using twice the thickness of conductors in the extension, the resistance comes out as about the same as the standard value.
The additional inductance of the extended cable is not mitigated by thicker wires, of course, but should not matter in the first place. There is adequated decoupling on all quality mother boards.
The 3V3 sense wire (light grey) typically found in ATX power supplies could not be connected to the 3V3 component power supply I used: surprisingly, there was no sense connection terminal. So I discarded the sense wire. Alternatives would be to use a different component supply or to modify the component supply (an easy modification in general, but always add a 100 ohm internal path to prevent overvolting if the sense wire goes open circuit for any reason).
Radiated EMI is a potential worry. The umbilical is normally inside a screened PC cabinet. Now it is external and twice as long. It would be fairly easy to constuct a foil screen around the umblical using cooking foil laced over with lacing cord and tinned copper. Copper braid is an (expensive) alternative. I did not bother with that.
Owing to the decoupling capacitors at both ends of the umbilical, the amount of high-frequency current in the conductors is fairly small. The radiated EMI will be proportional to the current being carried and the distance apart of the conductors. Since the current is balanced in the two directions, the radiation will tend to cancel and will be reduced in proprotion to how tightly the umbilical is bound and twisted. I used tight binding.
My second output to the USB hubs is much more of a concern for EMI. This creates a large triangle between the PC, the hub and the power supply. In retrospect, it is much better to take the hub supply from the PC end of the umbilicle, since this produces a linear arrangement instead of a closed triangle. I mitigated the problem with a large ferrite ring (see next section) on the hub supply. Well at least I assume this works, I've not made any measurements with an AM or FM radio or professional gear.
Earth loops cause problems for audio connections and for low frequency RF emissions (EMI/EMC). See separate note: Studio Wiring and Earth Loops.
It is common to use ferrite rings around cables to alleviate noise problems. Such a ring acts as a common-mode choke or ballun (balanced-to-unbalanced transformer). It essentially makes a break in the cable that still preserves the voltage difference between the conductors of the cable. This would be perfect to solve audio problems arising from earth loops in principle. But the inductance of commonly-used rings is insufficient to make any difference at mains frequencies where earth loops cause hum. Such rings only make a difference at radio frequencies. They help stop the cable from acting as a transmitting antenna for RF noise generated inside a device such as a lap top computer. To make a difference at mains frequencies, the ring would need to be sufficiently large for the audio cable to be looped through it many hundreds of times (I should do the calculation!)
A standard PC power supply joins mains earth to chassis earth and to the ground (OV) rail. A second connection to chassis and ground occurs with a metal motherboard enclosure where the shielded connectors mate with the motherboard backplate. This creates a small earth loop with two current paths between the PSU and the motherboard. A third and fourth path are made up from the CPU 12V supply cable and the SSD supply cable, back down the SATA screen. These paths can possibly degrade the performance of the on-board audio, which is single-ended. But most professional audio will instead use external sound cards or sound cards with differential (balanced) signals.
All the components I used were floating, so it was a simple matter to connect up all the chasis rails and metal panels and connect the combination to the mains protective earth at the IEC inlet connector. The metal grilles on the top cover are also earthed via copper contact wires (visible in the photos).
The connection between ground and earth is achieved by running an earth wire down the umbilical to the main 24-way molex connector to join one of the ground (black/0V) wires very close to the connector. This approach minimises the digital noise generated between ground and earth by the dynamic load current (as mentioned above).
Also, be careful with the grounding of the secondary of the linear supply component. The smoothing capacitors should be connected to the centre-tap of the transformer. A wire from this joint should then connect to the main ground. Any other arrangement, including a global star, where everything is joined together at one point, will be inferior. So do not install the wire marked 'no' at the point where it is shown on sheet 1. If you do not do this correctly, you will couple some of the smoothing current hum in to the rest of the design. Although there's actually likely be negligable current at this point with a typical PC, this is a good principle to follow for other applications.
Additionally, make sure the input-side 0.1 uF capacitor to the 7912 regulator is close to the regulator. The input terminal of these regulators can exhibit negative resistance which can cause oscillations if there is a lot of inductance from a long wire between the input capacitor and the regulator.
A wooden enclosure for a power supply is not ideal for two reasons. Firstly, it does not shield against radiated electromagnetic emissions (EMI). However, each component PSU is shielded. Moreover, conducted EMI are likely to greater than radiated and the radiation from the umbilical cable is likely to be the most severe radiation component. This is necessarily outside the box.
Secondly, a wooden enclosure does not behave well in an electrical fire. It combusts. Those who are worried by this should use fire-retarding paint or avoid wood. In my experience, fires from power supplies tend to be fairly short-lived: at worst, an individual component produces a small flame for tens of seconds. Given the distance of the wood from any of the power semi-conductors, the chance of the wood catching fire is very small indeed. In my case design, very little of the wood is horizontal and there is plenty of convection space. Heat during normal operation is not an issue.
A hybrid-cooled PC power supply uses both convection and forced air, switching the fan on when needed. This is also called a 'semi-passive' supply. Many off-the-shelf ATX power supplies control the voltage to the fan using a thermistor to get this effect.
Here we have presented a fully-passive design at the cost of using a couple of cubic feet in the office or studio. By suspending our design on hooks, behind the vanity board at the back of a desk that was pushed up against the wall, it utilises otherwise-wasted space and has access to an abundunt air supply.
If you (are mad enough to) make your own one of these, remember it is good practice to insulate all of the mains terminals inside the box so that it is relatively safe, even with the cover off. Also, keep the mains wires completely separate from the low voltage wires, bundling them separately. I included the chassis earth wire to each component PSU in the same mains bundle. Keeping the low and high voltage wires in separate bundels improves EMC and keeps them isolated should there be any melting of insulation from a high-current fault.
See also www.quietpc.com/powersupplies.
This material is provided in good faith, but you use it at your own risk. I will deny any liability for any losses or infringements that may arise.
END