How a Computer Power Supply Is Made: Topologies, 80 PLUS, Digital Control

A computer power supply is made by winding copper onto ferrite cores to build transformers and inductors, soldering those magnetics, power semiconductors and capacitors onto a circuit board, and wrapping the result in a fan cooled steel case. Electrically it does five jobs in order: it filters noise from the mains, rectifies AC to DC, corrects the power factor while boosting to a high DC bus, switches that bus through an isolating transformer down to 12 V, and derives 5 V and 3.3 V from the 12 V. Every unit is then hi-pot tested, load tested and often burned in, and the design is certified for safety, emissions and efficiency.

This post goes one level deeper than our walkthroughs of how a desktop PC is made and how a rack server is made. It covers the topologies inside a modern PSU, what the 80 PLUS levels actually measure, what changed with the ATX 3.x guidance and the 12V-2x6 connector, and how digital power supplies expose telemetry and accept firmware. It ends with read only commands for checking a PSU from Linux and from a server's management controller.

Power supply makers do not publish yields, burn-in durations or exact test recipes. Where no primary source gives a number we say so, and every arithmetic example is labelled as an illustration.

Key takeaways

• A switch mode PSU has a fixed chain of stages: EMI filter, bridge rectifier, active power factor correction (PFC) boost, a bulk capacitor, an isolated DC-DC converter (often an LLC resonant half bridge) with synchronous rectification, and small DC-DC converters that make 5 V and 3.3 V from the main 12 V.
• 80 PLUS is a certification run by CLEAResult. For 115 V desktop units it sets minimum efficiency at 20, 50 and 100 percent of rated load for Standard through Platinum, and adds a 10 percent point for Titanium and a 5 percent point for the newer Ruby level. Testing follows the Generalized Test Protocol, revision 6.7.3 dated 22 July 2025.
• ENERGY STAR Computers Version 9.0 took effect on 27 October 2025. It requires internal power supplies to meet efficiency minimums at 10, 20, 50 and 100 percent load and a power factor of at least 0.90 at 50 percent load for units of 75 W and above.
• Intel's ATX Version 3 Multi Rail Desktop Platform Power Supply Design Guide (revision 2.1a, November 2023) defines how far a PSU must survive power excursions, up to 200 percent of rating for 100 microseconds on units above 450 W with a 12V-2x6 connector. Revision 2.1 renamed the 12VHPWR connector to 12V-2x6 after the board header's internal pin lengths changed.
• The 12V-2x6 connector carries two sideband pins, SENSE0 and SENSE1, that tell the graphics card whether the PSU can supply 600, 450, 300 or 150 W through that cable.
• A digital PSU runs firmware on a microcontroller or digital signal controller and reports voltages, currents, power, temperatures, fan speed and fault status over PMBus. In servers the baseboard management controller reads those values and can update PSU firmware, which Redfish models with a FirmwareVersion property on each power supply and the UpdateService actions.
• Safety for IT power supplies is assessed against IEC 62368-1, whose current fourth edition was published on 26 May 2023. It classifies energy sources and prescribes safeguards against them rather than listing component rules.

Copper, ferrite and silicon to certified PSU: the step table

The table follows one PSU design from raw materials to a shipped, certified unit. Steps 1 to 4 happen at suppliers; steps 5 to 12 happen on the PSU maker's line; step 13 happens once per design (and again after significant changes) at accredited test laboratories and certification programs.

The production hi-pot voltages, burn-in times and sampling rates come from the safety standard, the certification body's factory follow-up program and the maker's quality plan. None of them is published in the free material we could fetch, so we do not quote numbers.

A dated timeline of the specifications

Every date below was checked against the primary document listed in the Sources section.

What does a PSU do, stage by stage?

A desktop ATX supply or a server's hot swap supply takes 100 to 240 V AC and turns it into tightly regulated low voltage DC. The modern design is a switch mode power supply: instead of a heavy 50 or 60 Hz transformer, it chops a high DC voltage at tens or hundreds of kilohertz, which lets a small ferrite transformer do the isolation. Follow the energy from the wall.

Stage 1: EMI filter and inrush limiting

The first parts after the AC inlet are a fuse, a varistor for surges, X capacitors across the line, Y capacitors from each line to earth, and one or two common mode chokes. Together they stop the switching noise generated inside the PSU from travelling back out onto the mains wiring, and they attenuate noise arriving from the mains. A thermistor or a resistor with a bypass relay limits the inrush current that would otherwise flow into the empty bulk capacitor at switch on. Intel's design guide lists input over current protection, inrush current and input under voltage behaviour as required items for desktop PSUs.

Stage 2: Bridge rectifier

Four diodes in a bridge turn AC into a rectified, pulsing DC. In high efficiency designs the diode bridge is a measurable loss, because two diodes conduct at all times and each drops close to a volt. Some top efficiency designs replace it with transistors in a "bridgeless" or "totem pole" PFC arrangement, which removes most of that drop.

Stage 3: Active power factor correction

A plain rectifier feeding a capacitor draws current only near the peaks of the voltage wave. That current is full of harmonics and gives a poor power factor. Equipment connected to public low voltage networks in many markets must keep harmonic currents under the limits in IEC 61000-3-2, which covers equipment drawing up to 16 A per phase. Efficiency programs ask for more: 80 PLUS requires a true power factor of at least 0.90 at full load, and ENERGY STAR Version 9.0 requires at least 0.90 at 50 percent load for internal supplies of 75 W and above.

The standard solution is an active PFC boost converter. A controller switches a MOSFET so that the current drawn through the PFC inductor follows the shape of the input voltage, while a boost diode charges a bulk capacitor to a regulated DC level well above the peak of the highest line voltage. A Texas Instruments application note on a 350 W interleaved PFC stage, for example, regulates the output at 385 V DC across an 85 to 265 V input range. Interleaving runs two boost stages 180 degrees out of phase, so their ripple currents partly cancel; the same note shows that this shrinks inductor size and output capacitor ripple current, with the best cancellation at a 50 percent duty cycle.

Stage 4: The bulk capacitor and hold-up time

The large aluminium electrolytic capacitor on the PFC output is the PSU's energy store. If the mains drops for a moment, this capacitor keeps the DC-DC stage running. Intel's guide (revision 2.1a, Table 4-8) requires a desktop PSU to hold its outputs in regulation for 12 ms at 100 percent load after a loss of input power at the low end of the nominal range, and recommends 17 ms at 80 percent load. It also requires the PSU to drop the PWR_OK signal before regulation is lost, so the motherboard can shut down in order.

Stage 5: The isolated DC-DC converter

The main converter takes roughly 380 to 400 V DC and produces 12 V across a transformer that provides safety isolation between mains and everything you can touch. Older and cheaper designs use a two transistor forward converter or an active clamp forward. Most current mid and high efficiency designs use an LLC resonant half bridge or full bridge, described in the next section.

Stage 6: Secondary rectification

The transformer's secondary output must be rectified again. Schottky diodes used to do this; now synchronous rectification uses MOSFETs switched in time with the transformer, because a low resistance MOSFET drops far less voltage than a diode at 12 V and tens of amps.

Stage 7: DC-DC converters for 5 V and 3.3 V

Older "group regulated" designs produced 12 V and 5 V from the same transformer and regulated them together, which led to poor cross regulation when one rail was loaded and the other was not. Modern desktop PSUs regulate only 12 V on the main transformer and use small non isolated buck converters, usually on a daughter board near the output, to produce 5 V and 3.3 V. A separate small flyback converter produces the +5 V standby rail (5VSB), which stays on whenever the PSU has mains power so the system can wake from sleep.

Stage 8: Supervision and protections

A supervisor circuit or the digital controller watches every output. It drives PWR_OK, which Intel defines as a TTL compatible signal asserted high only when the +12, +5 and +3.3 V outputs are within regulation and enough energy is stored to ride through the hold-up time, and it trips the protections described later.

Which topologies are used, and why?

Active PFC

Active PFC is now universal in desktop PSUs sold for markets with efficiency and harmonics rules. The boost controller has two control loops: a fast inner current loop that makes the inductor current follow a sinusoidal reference, and a slow outer voltage loop that holds the bulk capacitor at its target voltage. The output of a PFC stage is therefore not a perfect DC level; it has a ripple at twice the line frequency, which the next stage has to tolerate.

LLC resonant conversion

An LLC converter places a resonant tank, made of a series inductor, a series capacitor and the transformer's magnetising inductance, between a half bridge of MOSFETs and the transformer. The controller regulates the output by changing switching frequency rather than pulse width. Near the tank's resonant frequency the transistors switch while the voltage across them is close to zero (zero voltage switching), and the secondary rectifiers turn off with little reverse recovery. Switching loss falls sharply, which is why LLC sits inside most supplies that reach the higher 80 PLUS levels. The cost is design complexity: the operating frequency moves with load and input voltage, so the magnetics, the frequency range and light load behaviour (when many controllers switch to burst mode) all need care.

Synchronous rectification

At 12 V and 80 A, a diode with a 0.5 V forward drop would dissipate about 40 W on its own. A pair of modern MOSFETs with a few milliohms of on resistance dissipates a small fraction of that. The controller must switch them exactly in step with the transformer to avoid shoot through, which is one of the places where digital control helps.

DC-DC for the minor rails

Generating 5 V and 3.3 V with buck converters from the regulated 12 V means each rail holds its voltage regardless of how the others are loaded. Intel's guide sets voltage regulation limits, ripple and noise limits and a required power sequencing rule for the +5 V and +3.3 V outputs, which DC-DC daughter boards make easier to meet.

How are PSUs built and tested on the line?

Magnetics and boards

Magnetics are the most hand intensive part of a PSU. The main transformer is wound on a bobbin with primary and secondary windings separated by insulating tape layers or insulated wire grades, which give the creepage and clearance the safety standard requires between mains and the low voltage side. High frequency windings often use Litz wire, many fine insulated strands twisted together, because at hundreds of kilohertz current crowds to the surface of a solid conductor. The PFC inductor and output inductors carry large DC currents, so their cores are gapped to avoid saturation, and wound parts are commonly varnish impregnated to stop buzzing.

The main board mixes machine placed, reflow soldered surface mount parts with heavy through hole parts that are wave or selectively soldered. High current paths use thick copper or soldered busbars, and slots milled through the board under the isolation barrier lengthen the surface path between primary and secondary.

Hi-pot, ground bond and burn-in

Every finished unit gets an electric strength (hi-pot) test between the mains side and the touchable low voltage side and between mains and earth, plus a ground bond test of the protective earth path. Test voltages and durations come from the safety standard and the certification body's factory program, which we could not read in a free primary source, so we do not quote them. Burn-in then runs units under load, often in a heated chamber, to catch early life failures in capacitors, solder joints and semiconductors; how long, how hot and at what load are internal decisions makers do not publish.

Functional load testing

Automated test equipment applies programmed electronic loads to each rail and checks the unit against its specification. Intel's guide gives a sense of what is checked: DC voltage regulation per rail, output ripple and noise, capacitive load stability, closed loop stability, power sequencing, turn on time, rise time, overshoot at turn on and off, PWR_OK timing, hold-up time, and each protection trip. For ATX 3.x units the excursion test described later adds a dynamic load profile.

Efficiency, safety and EMC certification of the design then happen on samples at outside laboratories, as step 13 of the table describes. Intel's guide lists EMC emissions, immunity, magnetic leakage field and voltage fluctuation and flicker as required items, alongside North American and international safety approvals.

What do 80 PLUS and other efficiency certifications measure?

The 80 PLUS levels

80 PLUS is a performance specification and certification program for internal power supplies, run by CLEAResult. Its base requirement is at least 80 percent efficiency at 10, 20, 50 and 100 percent of rated load with a true power factor of 0.9 or higher, and it offers seven levels: Standard, Bronze, Silver, Gold, Platinum, Titanium and Ruby. The program lists seven categories of certified supplies: 115 V internal desktop, 230 V EU internal desktop, 115 V industrial, 230 V internal redundant AC data center, 380 V internal redundant DC data center, 277 V internal redundant AC data center and 480 V internal redundant AC data center. Its ratings page also lists 800 V DC for redundant supplies. Testing and certification cost between 3,500 and 8,000 US dollars per product, according to the program.

The table below copies the minimum efficiency requirements from the 80 PLUS ratings page for the two desktop categories.

For redundant data center supplies at 230 V and 277 V the bars are higher at mid load: Gold requires 85, 88, 92 and 88 percent at 10, 20, 50 and 100 percent load, Titanium requires 90, 94, 96 and 91 percent, and Ruby requires 90 percent at 5 percent load, 91 at 10, 95 at 20, 96.5 at 50 and 92 at 100. For 380 V DC, 800 V DC and 400 and 480 V AC redundant supplies, Ruby asks for 97 percent at 50 percent load.

Three things in that table are easy to miss. First, a desktop Gold certificate says nothing about efficiency at 10 percent load, where many PCs spend most of their time idling. Only Titanium and Ruby add a 10 percent point. Second, efficiency peaks around half load and falls at both ends, which is why the 100 percent figure is lower than the 50 percent figure for every level. Third, the 230 V EU numbers are higher for the same level because a higher input voltage means lower input current and lower conduction losses in the front end.

The power factor requirement differs by level. The ratings page lists a power factor of at least 0.90 for the lower levels and at least 0.95 for Platinum, Titanium and Ruby in the 115 V table, but the per level summary cards on the same page show 0.95 for Bronze, Silver and Gold at 115 V. The tables and the cards disagree, so check the certificate and test report for the exact power factor a given unit was held to.

How efficiency is tested

The 80 PLUS method is the Generalized Test Protocol for Calculating the Energy Efficiency of Internal AC-DC and DC-DC Power Supplies, prepared by EPRI and CLEAResult. Revision 6.7.3 is dated 22 July 2025. The main rules are:

• Input. The unit is fed from an AC reference source at the specified voltage within 1 percent and the specified frequency within 1 percent, at either 115 V 60 Hz or 230 V 60 Hz, whichever is closest to the nameplate. The source's voltage distortion must stay at or below 2 percent up to the 13th harmonic, following IEC 62301.
• Load points. Efficiency is measured at 10, 20, 50 and 100 percent of rated nameplate output power. Lower load points may be recorded for supplies typically used in redundant configurations. The standby rail is loaded like any other rail, and a rail's manufacturer specified minimum current overrides the calculated value when it is higher.
• Steady state. The unit runs at each load point for at least 15 minutes, and a steady state is reached when total input power over two consecutive five minute intervals changes by no more than 1 percent. The warm-up section adds a tighter recommendation of 0.2 percent.
• Environment. Ambient temperature is held at 23 degrees Celsius, plus or minus 5 degrees.
• Fans. If a fan cycles on and off with temperature, input and output power are integrated over 30 minutes after thermal equilibrium or over five fan cycles, whichever comes first.
• Instruments. Power of 0.5 W or more is measured with an uncertainty of 0.5 percent or better at 95 percent confidence, with a line impedance stabilisation network recommended to keep light load power factor readings stable.

Efficiency is simply DC output power divided by AC input power at each load point. The test report also records power factor and current distortion.

ENERGY STAR and the Intel design guide

ENERGY STAR does not certify bare PSUs; it certifies computers, and its Computers Version 9.0 specification, effective 27 October 2025, sets requirements for the internal power supply inside them. Supplies must be tested with the Generalized Test Protocol revision 6.7.2 and meet:

Supplies under 75 W have to meet only the efficiency minimums. The specification also says test reports from testing entities recognised by EPA are accepted, which is why the same 80 PLUS lab data often supports both.

Intel's design guide sets a much lower floor for any compliant desktop PSU, 70 percent at full load, 72 percent at typical (50 percent) load and 65 percent at light (20 percent) load at 115 V, and adds a low load requirement: at 10 W or 2 percent of rating, whichever applies, at least 60 percent efficiency is required and 70 percent recommended. Its reasoning is that computer idle power has fallen so far since 2005 that PSU loss at idle has become a large part of total AC draw.

What changed with ATX 3.x and the 12V-2x6 connector, and why do GPU transients matter?

Power excursions became a specification item

Before PCI Express CEM revision 5.0, add-in card power was confined to hard limits such as 10, 75, 150, 225 or 300 W, according to Intel's guide. Many cards exceeded those limits for short periods anyway, and power supplies were never designed for it, so brief spikes could trip a PSU's over current protection or cause voltage droop, especially with more than one card installed. The CEM 5.1 specification introduced Sustained Power, defined as the average power through a single power cable in a one second moving window, and allowed cards to exceed their nominal limit briefly as long as the time average stays inside it. Excursions are allowed through the 12V-2x6 and 48VHPWR connectors and the slot's 12 V, but not through the legacy 6 pin and 8 pin connectors and not on the 3.3 V domains.

What the PSU must survive

Intel's design guide converts those card allowances into PSU requirements. Table 3-3 of revision 2.1a:

The test applies a repeating load: a peak for the excursion time, then a lower power for the rest of the period, chosen so that the RMS power equals the PSU's rated wattage. For a 1000 W unit Intel's Table 3-4 gives 2000 W for 100 microseconds followed by 917.7 W for 1900 microseconds, 1800 W for 1 ms with 897.3 W for 11.5 ms, 1600 W for 10 ms with 881.6 W for 70 ms, and 1200 W for 100 ms with 923.8 W for 300 ms. You can check the first row: the square root of (0.05 times 2000 squared plus 0.95 times 917.7 squared) is about 1000 W.

Why it matters: a GPU's power controller reacts to workload changes in microseconds. If the PSU's over current protection is set tight, or its output capacitance and control loop cannot hold 12 V during a fast step, the PC reboots or shuts down under a heavy game or training load even though the average power is well inside the label rating. The excursion requirement turns "the PSU should cope" into a test.

The 12V-2x6 connector and its sense pins

The 12VHPWR connector introduced for high power cards was renamed 12V-2x6 in revision 2.1 of Intel's guide (September 2023) because the board header's internal pin lengths were modified. The cable plug side did not change and remains compatible. The connector carries 12 power contacts in two rows of six, plus four small sideband contacts.

Two of the sideband contacts, SENSE0 and SENSE1, are required on both the PSU and the card. The PSU configures them by grounding or leaving each open to advertise 600, 450 or 300 W of sustained power through that cable, and by shorting the two together (while leaving them floating with respect to ground) to advertise 150 W. Intel's guide points out the failure mode: if the 150 W short were also tied to ground, the card would read it as 600 W. The two optional sideband signals are CARD_PWR_STABLE and CARD_CBL_PRES#. The advertised level is the maximum sustained power through that connector alone, separate from the 75 W class power a card can take from the slot.

Revision 2.1a added guidance on 12V-2x6 cable plug connection, including notes on temperature rise. The practical lesson, which the guide's OCP section echoes, is that connectors and cables must not melt or be damaged before the over current protection trips. Full insertion of the plug, avoiding sharp bends close to the connector, and not mixing cables between PSU models all follow from that.

Hold-up time and timing

Revision 2.1 also split hold-up time into a required level of 12 ms at 100 percent load and a recommended 17 ms at 80 percent load; before that only one level was required. Revision 2.0 had already reclassified timing values so that previous recommended timings became required and the previous required values became legacy timings, and it asked that every PSU label state which revision of the specification it meets.

What is a digital PSU, and how is server PSU firmware updated?

Analog versus digital control

In an analog PSU the control loops are built from error amplifiers, comparators and dedicated controller chips; trip points are set by resistor dividers. A digital PSU replaces some or all of that with a microcontroller or digital signal controller that samples voltages and currents, runs the control loops in firmware, and exposes what it sees over a management bus. Many designs are hybrids: an analog PFC controller and a digital LLC and secondary side controller, or analog loops supervised by a microcontroller that handles fan control, protections and telemetry.

PMBus

PMBus is an open command language for power conversion devices that runs over SMBus, a variant of I2C. The current specification is revision 1.4. Revisions 1.3 through 1.4.1 were split into three parts: general requirements and transport, the command language, and the Adaptive Voltage Scaling Bus (AVSBus), which has since become its own specification. Revision 1.0 Part I is dated 28 March 2005 and revision 1.3 is dated March 2014.

The command set lets a host read input and output voltage, current and power (for example the commands the Linux kernel maps as READ_VIN, READ_VOUT, READ_IIN, READ_IOUT, READ_PIN and READ_POUT), read temperatures and fan speeds, read status registers, and read manufacturer limits such as MFR_VIN_MAX or MFR_POUT_MAX. A PAGE command selects which output a multi rail device is talking about. A WRITE_PROTECT command controls whether settings can be changed over the bus, and the Linux PMBus core has flags for devices whose write protection works differently.

How Linux exposes PMBus telemetry

On Linux, PMBus devices appear through the hwmon subsystem. The kernel's generic pmbus driver does not probe for devices on its own, because there is no register that reliably identifies a PMBus device, so devices are declared in a device tree, by the platform, or instantiated by hand. Once bound, each sensor appears as a sysfs file: inX_input for measured voltage from READ_VIN or READ_VOUT, currX_input for current, powerX_input for power from READ_PIN or READ_POUT, tempX_input for temperature and fanX_input for fan speed. Limit attributes are read-write; measurement attributes are read only. The hwmon interface reports power in microwatts.

How servers manage PSUs

In a rack server, the PSUs sit on a PMBus segment owned by the baseboard management controller (BMC). The BMC polls them, raises alerts on faults, balances or puts redundant supplies into standby, and reports what it sees through IPMI sensors and Redfish.

DMTF's Redfish PowerSupply schema (v1_6_0, release 2024.1) describes each unit at a URI of the form /redfish/v1/Chassis/CHASSIS_ID/PowerSubsystem/PowerSupplies/PSU_ID, with the chassis and supply identifiers filled in. Its properties include Manufacturer, Model, PartNumber, SerialNumber, FirmwareVersion, PowerCapacityWatts, InputRanges, LineInputStatus, EfficiencyRatings (pairs of LoadPercent and EfficiencyPercent), HotPluggable, ProductionDate and a link to a metrics resource. Version 1_6_0 added Certificates, a link to certificates for device identity and attestation. The schema also defines a Reset action that resets the power supply.

Firmware updates go through the Redfish UpdateService. Its SimpleUpdate action updates software components from an image URI, with an option to verify the remote server's certificate, and MultipartHttpPushUri accepts a multipart HTTP or HTTPS push defined by the Redfish specification. The older HttpPushUri message format is vendor specific. Behind the BMC, the actual transfer of an image into a PSU's controller uses the PSU maker's own bootloader and commands, which are not public in general.

Why PSU firmware is a security and reliability concern

NIST SP 800-193, Platform Firmware Resiliency Guidelines (May 2018), frames firmware security as three jobs: protect firmware from unauthorised changes, detect unauthorised changes that happen, and recover from them rapidly and securely. A PSU controller is a small platform component with its own firmware, so the same questions apply on a smaller scale. Does the PSU accept only signed images? Can the BMC read back the running version? If an update fails halfway, does the unit fall back to a known good image? Is PMBus write access limited to the BMC? A PSU that accepts unauthenticated changes to its protection limits is a safety problem as well as a security one, which is why write protection on the management bus and controlled BMC access both matter. We do not know of a public, primary source describing a PSU firmware attack in the wild, so treat this as defensive design guidance rather than a threat report.

What safety approvals apply, and what do OCP, OVP, OTP and SCP do?

IEC 62368-1

IEC 62368-1 covers the safety of audio, video, information and communication technology equipment with rated voltage up to 600 V. It is a hazard based standard: it classifies energy sources (electrical, thermal, mechanical, chemical, radiation) and prescribes safeguards against each, rather than listing component specific rules. The IEC says it does not cover performance or functional characteristics. The third edition was published on 4 October 2018 and withdrawn on 26 May 2023, when the fourth edition was published. National and regional adoptions carry the same content under local numbers.

For a PSU this standard drives the isolation design (insulation grades, creepage and clearance, the transformer construction), touch current limits, fire enclosure and component temperature limits, and the abnormal and fault condition tests. In North America, certification is done by accredited third party laboratories, and the mark on the label names the laboratory that certified it. Intel's guide lists both North American and international safety approvals as required, with catastrophic failure protection recommended.

The output protections

Server PSUs add input side protections, current sharing between redundant units, and fault logging that the BMC can read.

Worked examples with numbers

These are illustrations built from the published minimums. Real units usually beat the minimum; read the test report for the measured values.

Example 1: wall power and heat at a given DC load

A PC draws 500 W DC from a 1,000 W PSU, so the PSU sits at 50 percent load.

• With a 115 V 80 PLUS Gold minimum of 90 percent at 50 percent load, AC input is 500 divided by 0.90, which is 555.6 W. The PSU turns 55.6 W into heat.
• The same unit on 230 V EU mains would need only the 230 V Gold minimum of 92 percent: 500 divided by 0.92 is 543.5 W, a loss of 43.5 W.
• A 115 V Titanium unit at its 94 percent minimum draws 500 divided by 0.94, or 531.9 W, losing 31.9 W.

The difference between the Gold and Titanium minimums is 23.7 W. At eight hours a day for a year, that is 23.7 times 8 times 365, or about 69.2 kWh. At an assumed tariff of 0.17 US dollars per kWh, it is about 11.76 dollars a year.

The input current follows from power factor. With 555.6 W of real power at a power factor of 0.95, apparent power is 555.6 divided by 0.95, which is 584.8 VA, and at 115 V the current is 584.8 divided by 115, about 5.1 A.

Now take the same PC idling at 100 W DC, 10 percent of a 1,000 W rating. The desktop Gold level has no 10 percent requirement at all, while Titanium requires 90 percent there. At 90 percent the input is 111.1 W. If a non Titanium unit only managed 80 percent at that point, the input would be 125 W, a 13.9 W difference that runs every hour the machine idles. That is why oversizing a PSU for a machine that idles most of the day can cost more energy than it saves.

Example 2: sizing a PSU for a GPU build

Assume a build with a graphics card rated for 450 W sustained through a 12V-2x6 cable, a CPU you have limited to 200 W, and 100 W for the rest of the platform (motherboard, memory, storage, fans, USB). Sustained demand is 450 plus 200 plus 100, or 750 W.

• An 850 W ATX 3.x unit carries 750 W at about 88 percent load. That is legal but leaves the PSU near the top of its efficiency curve's falling side and leaves little thermal margin.
• A 1,000 W unit carries it at 75 percent load. Under Intel's Table 3-3 a 1,000 W unit with a 12V-2x6 connector must survive 2,000 W for 100 microseconds, 1,800 W for 1 ms, 1,600 W for 10 ms and 1,200 W for 100 ms.
• Suppose, for illustration, that a workload change makes the card briefly draw 900 W for under a millisecond while the CPU and platform stay at 300 W. The momentary total is 1,200 W. That is inside the 1,800 W the 1,000 W unit must handle for 1 ms and inside the 1,530 W (180 percent) an 850 W unit must handle for the same time. Either should hold; the larger unit has more headroom if a second card or a higher CPU limit is added later.

Check the 12V-2x6 rating too: 450 W at 12 V is 37.5 A through the connector's six 12 V contacts, about 6.25 A per contact, compared with 50 A in total (about 8.3 A per contact) at the 600 W setting. The cable and connector must be rated for the setting the PSU advertises, which is why the sense pin configuration matters.

Then look at efficiency at idle. If the same build idles at 90 W most of the day, a 1,000 W unit sits at 9 percent load, below the 10 percent point most certifications test. A Titanium or Ruby certificate is the only place on the 80 PLUS label that speaks to that region.

Example 3: sizing the bulk capacitor for hold-up

A 1,000 W PSU must keep its outputs in regulation for 12 ms at full load after the mains disappears. Assume the DC-DC stage after the bulk capacitor is 96 percent efficient, the capacitor starts at 390 V and the LLC stage can regulate down to 300 V.

• Energy needed: 1,000 divided by 0.96 is 1,041.7 W, times 0.012 s, is 12.5 J.
• Usable capacitor energy is one half times C times (390 squared minus 300 squared), which is one half times C times 62,100.
• Solving, C is 2 times 12.5 divided by 62,100, about 403 microfarads, before allowing for capacitor tolerance and ageing.

Meeting the recommended 17 ms at 80 percent load takes 800 divided by 0.96 times 0.017, about 14.2 J, which needs about 457 microfarads under the same assumptions. This is why high wattage PSUs carry one or two very large capacitors, and why a design that lets the LLC stage regulate down to a lower bus voltage can use a smaller one.

Step by step: read a PSU's identity, telemetry and firmware version

These steps are read only. They work on a Linux desktop or workstation that exposes PSU telemetry, and on a server with a BMC that supports IPMI or Redfish. Do not write to PMBus registers or run firmware updates unless your platform vendor documents the procedure for your exact model.

1. Find hwmon devices on a Linux host

# List every hwmon device and the driver that registered it
for d in /sys/class/hwmon/hwmon*; do
  printf '%s  %s\n' "$d" "$(cat "$d/name" 2>/dev/null)"
done

# If lm-sensors is installed, print a readable summary
sensors

Most desktop PSUs expose nothing here, because they have no digital interface. Some premium desktop PSUs report over USB to the maker's software, and Linux has drivers for some of them that appear as hwmon devices. Servers and boards with PMBus devices declared in the device tree or platform code show entries named after the PMBus driver.

2. Read PMBus values from sysfs

# Replace hwmon3 with the device you found in step 1
cat /sys/class/hwmon/hwmon3/name
cat /sys/class/hwmon/hwmon3/in1_input      # often input voltage, millivolts
cat /sys/class/hwmon/hwmon3/power1_input   # input power, microwatts
cat /sys/class/hwmon/hwmon3/temp1_input    # temperature, millidegrees Celsius
ls /sys/class/hwmon/hwmon3/                # see which attributes exist

Label files such as in1_label and power1_label say which reading is input and which is output. Do not write to the _max, _crit or _cap files: they are limits, and on many devices they change real protection behaviour.

A caution on bus scanning. Tools that probe every address on an I2C or SMBus segment can upset some devices, including power controllers. Listing buses with i2cdetect -l is safe; probing a bus with i2cdetect -y is not something to do on a production server.

3. Read PSU sensors and inventory through IPMI

# Sensors of type Power Supply (presence, failure, input lost)
ipmitool sdr type "Power Supply"

# All sensors, filtered for PSU readings such as input power and temperature
ipmitool sensor | grep -i -E 'psu|ps[0-9]|power supply'

# Field replaceable unit records, often including PSU model and serial
ipmitool fru print

Run these on the host with the local IPMI driver loaded, or add -I lanplus -H BMC_HOST -U BMC_USER to query the BMC over the network. Prefer a read only BMC account for monitoring.

4. Read power supplies through Redfish

BMC=https://bmc.example.internal
# List the chassis, then the power supplies under the chassis you want
curl -s -u monitor "$BMC/redfish/v1/Chassis" | jq '.Members'
curl -s -u monitor "$BMC/redfish/v1/Chassis/1/PowerSubsystem/PowerSupplies" | jq '.Members'

# Read one supply: identity, firmware version, capacity and efficiency ratings
curl -s -u monitor "$BMC/redfish/v1/Chassis/1/PowerSubsystem/PowerSupplies/0" \
  | jq '{Manufacturer, Model, SerialNumber, FirmwareVersion, PowerCapacityWatts, EfficiencyRatings, Status}'

# Read live metrics through the link the supply provides
curl -s -u monitor "$BMC/redfish/v1/Chassis/1/PowerSubsystem/PowerSupplies/0/Metrics" | jq

Chassis and supply IDs vary by vendor; follow the Members links rather than guessing. Keep the BMC's TLS certificate verification on in production by installing its CA in your trust store rather than passing -k.

5. Plan a server PSU firmware update

Warning: a PSU firmware update changes the controller that protects your server. Update one supply at a time, only in redundant configurations, and only with the image and procedure your server vendor publishes for your exact PSU part number.

# Before: record every supply's model, part number and firmware version
curl -s -u monitor "$BMC/redfish/v1/Chassis/1/PowerSubsystem/PowerSupplies/0" \
  | jq '{Model, PartNumber, FirmwareVersion, Status}'

# Check the update service and which push methods it supports
curl -s -u monitor "$BMC/redfish/v1/UpdateService" \
  | jq '{ServiceEnabled, HttpPushUri, MultipartHttpPushUri, Actions}'

Confirm that the redundant supply reports "Health": "OK" and that the system's load fits on one supply before you start. Run the update itself with an administrator account through the vendor's documented SimpleUpdate or multipart push, targeting only the supply you mean to update. Afterwards, read FirmwareVersion again for each supply and check the BMC event log before moving to the next one.

6. Measure real efficiency at the wall

For a desktop, a plug-in power meter or a metered PDU outlet gives AC input power. Divide the DC power you can measure or estimate by the AC reading to get a rough efficiency figure, keeping in mind that inexpensive meters have their own error at low loads. For servers, compare the BMC's input power with its output power reading; Redfish and IPMI usually expose both.

Common mistakes

• Reading 80 PLUS Gold as "90 percent efficient". At 115 V the Gold minimum is 90 percent only at 50 percent load. At 20 and 100 percent it is 87 percent, and at 10 percent load there is no requirement.
• Oversizing by habit. A 1,600 W unit in a PC that idles at 80 W spends most of its life at 5 percent load, below every certified point except Ruby's 5 percent bar.
• Undersizing by average. Average power is not the right figure for modern GPUs. Check that the PSU meets the ATX 3.x excursion requirements, which a label or the maker's documentation should state.
• Mixing modular cables. Pinouts on the PSU side of modular cables are not standardised between makers or even between product lines. A cable from another PSU can put 12 V on a ground or sense contact.
• Partly seated 12V-2x6 plugs and sharp bends. The connector's current per contact is high. Push the plug fully home and avoid bending the cable hard right at the connector.
• Assuming a server PSU is just a box. Redundant supplies run firmware, report telemetry and can be updated. Leaving them out of the firmware inventory leaves out a controller sitting on the platform's management bus.
• Scanning SMBus on a production server. Address probing can put some power controllers into unexpected states. Use the BMC or the declared hwmon devices.
• Trusting telemetry as calibrated measurement. PMBus readings are good for trends and alarms. For efficiency claims, use a calibrated power analyser as the 80 PLUS protocol requires.
• Confusing safety marks with efficiency certificates. A safety certification says the unit's design was evaluated against IEC 62368-1 or its national adoption. It says nothing about efficiency, noise or transient response.

FAQ

What does 80 PLUS Gold mean?

For a 115 V desktop PSU it means the unit was tested to be at least 87 percent efficient at 20 percent load, 90 percent at 50 percent load and 87 percent at full load, under the 80 PLUS test protocol. The 230 V EU Gold level asks for 90, 92 and 89 percent at the same points.

What is the difference between 80 PLUS Titanium and Ruby?

Titanium adds a 10 percent load point (90 percent efficiency at 115 V) to the usual 20, 50 and 100 percent points. Ruby raises every point by about one percentage point and adds a 5 percent load point at 85 percent efficiency for desktop units.

Is ATX 3.1 the same as ATX 3.0?

They are successive revisions of Intel's desktop PSU design guidance. Revision 2.0 of the combined design guide (February 2022) added detail for the new 12VHPWR connector. Revision 2.1 (September 2023) renamed the connector 12V-2x6, added 150 W and 0 W sense encodings and changed hold-up time to a 12 ms required and 17 ms recommended pair. The current revision, 2.1a, carries ATX12V guidelines version 3.1.

Can I use a 12VHPWR cable with a 12V-2x6 graphics card?

Intel's guide states that the cable plug side did not change when 12VHPWR became 12V-2x6 and remains compatible with the new board header. The change was to the header's internal pin lengths. Use the cable that came with your PSU, make sure it is fully seated, and follow the PSU and card makers' instructions.

How many watts of PSU do I need for my GPU?

Add the GPU's sustained power, your CPU's configured power limit and roughly 75 to 150 W for the rest of the platform, then choose a unit that runs that total at around 50 to 80 percent load and meets the ATX 3.x excursion requirements. For a 450 W card, a 200 W CPU and 100 W of platform, a 1,000 W unit puts sustained load near 75 percent.

What is PMBus used for?

PMBus is an open command language for power devices over SMBus. It lets a host or BMC read voltages, currents, power, temperatures, fan speeds and status, and on some devices change settings, from power supplies and voltage regulators.

How do you update the firmware on a server power supply?

Through the server's BMC, using the vendor's published image and procedure, usually exposed through Redfish UpdateService actions or a vendor tool. Update one supply at a time in a redundant pair, confirm the remaining supply is healthy first, and read back FirmwareVersion afterwards.

What safety standard applies to PC power supplies?

IEC 62368-1, the hazard based safety standard for audio, video and information technology equipment, or its national adoptions. The current fourth edition was published on 26 May 2023.

Where Safeguard fits, honestly

Safeguard does not design, test, certify or analyse power supplies. It does not scan PSU or BMC firmware images, does not generate a hardware bill of materials, does not manage devices and does not measure power. Its models are single purpose security models for software. The fit is narrower and indirect: the software around power supplies is ordinary software. The telemetry exporters that read hwmon files or Redfish endpoints, the Python and Go tools that poll BMCs, the dashboards that graph input power, the automation that schedules PSU firmware updates and the container images that package all of it carry open source dependencies with their own known vulnerabilities. Safeguard scans those repositories, container images and SBOM files, shows which projects carry a given component and version, matches them against vulnerability databases with reachability analysis, evaluates your policies and gates, and can open fix pull requests when a fixed version exists, with autonomous remediation switched off by default.

Fix it without waiting for upstream

A PSU has no dependencies to update, but the systems around it do: every rebuilt telemetry collector or BMC automation image pulls its libraries again, and BMC firmware and management agents ship open source components of their own. The power hardware stays with its makers and your server vendor; the software you build around it is where supply chain hygiene still applies.

Whatever the vulnerability, most of the risk sits in two gaps. The first is the time between a fix being published and that fix actually running in your production builds. The second, and worse, is the time before any fix exists at all, while a maintainer works out a patch and a release schedule. Waiting through both gaps is a habit, not a requirement. Safeguard gives teams three ways to close them.

1. Autonomous remediation: pull requests, not tickets

When a fixed version exists, Safeguard's Griffin models open a tested, policy gated pull request that moves the affected dependency to the safe version. You can run it on one repository or in bulk across every project that carries the vulnerable version. Remediation covers eight categories, including dependencies, licenses, security posture and findings. Reachability analysis ranks the vulnerable code paths your application actually calls first, so the first pull requests you merge are the ones that matter most. Autonomous remediation ships switched off, with every category set to Manual, so your team decides how much runs on its own and what still needs a human approval.

2. Gold Open Source: the safe version in seconds, free

gold.safeguard.sh is free and needs no login. Search any CVE or package to see affected and fixed versions, CVSS, EPSS, CISA KEV status and a recommended safe version, or use the "Am I affected?" checker. The same data is available through a free read API, README status badges and RSS threat feeds. The safeguard-sh/gold-scan-action GitHub Action can fail a build on a known exploited vulnerability, on any known vulnerability, or never, so a version you have just removed cannot quietly come back in next week's merge.

3. Gold Registry: start from components that are already fixed

registry.safeguard.sh provides hardened, drop in replacements for upstream packages and container images. Every Gold artifact is held continuously to zero exploitable critical or high CVEs, zero malware and a completed license check, with SLSA L3 or higher provenance, a Sigstore signature and a public evidence page you can hand to an auditor. An artifact loses its Gold badge if it regresses. Teams can point their own package directory at the registry so hardened equivalents resolve without changing how developers install anything.

When upstream has no fix yet, or will not ship one, Safeguard first works to get the fix accepted upstream. If the issue is severe or upstream will not move, Safeguard builds a fork, fixes the vulnerability there and publishes a hardened Gold build, so you are not stuck on a maintainer's release schedule. You can also request a hardened, compatibility tested build of a specific package. This fork and patch flow is in early access today, so ask which packages it covers before you plan around it.

One honest limit: when the vulnerable code lives in a vendor appliance, a hosted service or an operating system runtime you do not build yourself, the fix still has to come from that vendor or distribution. The steps above apply to the open source dependencies and container images you build and ship, which is where most teams carry the largest and least tracked share of their exposure.

See how Safeguard would help

The first demo is a real Safeguard CLI session asked to fix a lodash CVE in a repository. It reads the finding and the advisory, asks before each command, edits package.json to lodash 4.17.21, runs npm install and npm test with your approval, prints a summary and pushes nothing. The same flow applies to a power telemetry dashboard or a BMC polling service your team maintains.

Click through it at your own pace, or open it full screen.

The second demo sets up the Safeguard runner on your own machine, approves it with its device code and then scans from the terminal. That suits infrastructure teams who build their monitoring and automation images on internal hardware and want scans to run there rather than in a hosted service.

Click through it at your own pace, or open it full screen.

The third demo connects Claude to the Safeguard MCP server with OAuth and shows how to choose which tools Claude can use. A data center engineer can then ask about the components in the repositories that hold their fleet tooling, without granting more access than they intend.

Click through it at your own pace, or open it full screen.

To try it on your own repositories, download Safeguard and follow the CLI documentation.

Quick check

Sources

• CLEAResult, 80 PLUS program overview
• CLEAResult, 80 PLUS program details
• CLEAResult, 80 PLUS PSU ratings explained
• CLEAResult, 80 PLUS certified 115 V internal manufacturers list
• EPRI and CLEAResult, Generalized Test Protocol for Calculating the Energy Efficiency of Internal AC-DC and DC-DC Power Supplies, Revision 6.7.3 (22 July 2025)
• Intel, ATX Version 3 Multi Rail Desktop Platform Power Supply Design Guide, document 336521 revision 2.1a
• Intel design guide 336521, PCIe add-in card power excursions
• Intel design guide 336521, PSU power excursion
• Intel design guide 336521, PCIe add-in card and PSU power budgets
• Intel design guide 336521, SENSE0 and SENSE1
• Intel design guide 336521, overall power supply efficiency levels
• Intel design guide 336521, voltage hold-up time
• Intel design guide 336521, PWR_OK
• Intel design guide 336521, over voltage protection
• Intel design guide 336521, over current protection
• Intel design guide 336521, short circuit protection
• Intel design guide 336521, over temperature protection
• Intel design guide 336521, revision history
• US EPA, ENERGY STAR Computers partner resources
• US EPA, ENERGY STAR Computers Version 9.0 Final Specification
• IEC, IEC 62368-1:2018 (third edition)
• IEC, IEC 62368-1:2023 (fourth edition)
• IEC, IEC 61000-3-2 harmonic current emission limits
• PMBus, home page
• PMBus, specification archives
• Linux kernel documentation, PMBus driver
• Linux kernel documentation, PMBus core driver
• Linux kernel documentation, hwmon sysfs interface
• DMTF, Redfish PowerSupply schema index
• DMTF, Redfish PowerSupply schema v1_6_0
• DMTF, Redfish UpdateService schema v1_14_0
• NIST, SP 800-193 Platform Firmware Resiliency Guidelines (May 2018)
• Open Compute Project, Rack and Power project
• Texas Instruments, SLUA369C, 350-W Two-Phase Interleaved PFC Pre-Regulator Design Review

Last reviewed: 24 October 2026.