PC and AI Rig Power Supply Calculator | PSU Sizing

The calculation uses Intel ATX 3.1 Design Guide, revision 2.1 (September 2023), checked against the public revision 2.1a PSU excursion section (November 2023). No paid PCI-SIG specification is used.

Continuous sizing. Sum CPU PSU-side watts, GPU board watts times GPU count and other DC loads. Multiply by one plus the entered headroom fraction. Headroom is a user planning choice. Supply rating is DC output; efficiency does not reduce the usable nameplate watts.

Peak sizing. Intel p.24 gives a total-card upper envelope above 75 W: the ratio is capped at 3 for intervals up to 100 µs, and follows $4-0.2171\ln(T_{\mu s})$ for the intervals evaluated here. Custom ratios may replace it. For each checkpoint, add all simultaneous GPU demands to the entered coincident CPU peak and remaining platform power. Intel p.27 supplies the whole-PSU excursion capability factors; these are different from GPU ratios. The minimum rating is the larger of the continuous budget and each peak divided by its PSU factor. The solver explicitly handles the 450 W branch boundary before rounding upward. Credit requires verified excursion capability; unknown PSUs receive no overload credit. Chart lines only join checkpoints.

Wall energy. Linear interpolation between CLEAResult desktop non-redundant certification anchors gives an estimate. Uncertified light-load points use the entered 5% assumption; below 5%, an assumed fixed conversion loss preserves that anchor. Titanium and Ruby use their published low-load anchors. Measured mode uses a constant efficiency. Wall watts equal DC watts divided by efficiency. Sum active and idle kWh over the entered operating days, then multiply by the entered price. Zero DC is treated as off, excluding standby. All cost and energy results are per rig.

Circuit screen. Compute active and idle true RMS current per rig as $P/(V_{ac}PF)$, using the same entered power factor for both. Use the larger current for every identical rig, add other shared current, and subtract from breaker rating times the entered continuous allowance. Both modeled states are screened regardless of the hours entered for annual energy. Use a power factor representative of both states. The allowance is user supplied, not a jurisdiction-specific compliance determination.

Connector network, derived here from Kirchhoff and Joule laws. Auxiliary load per connector is (GPU board watts minus slot watts) divided by equal connector count. At entered delivered voltage, $I=P/V$. In each six-way bank, branch resistance is mating-contact resistance plus common wire/other-path resistance. Then $I_j=I(1/R_j)/\sum_k(1/R_k)$ and mating-contact heat is $I_j^2R_{contact,j}$. Supply and return banks are solved independently: no artificial supply/return pin pairing. Their equivalent resistances give total voltage drop. Connector losses are displayed separately and are not added to component watts.

The Intel RMS laboratory recovery-power step validates the method against published examples. Its duty fractions are test conditions, not PC operating duty or electricity-cost assumptions.

• User wattages represent simultaneous load at the stated operating condition; CPU watts include VRM losses. Identical GPUs and identical rigs are assumed. CPU and non-GPU powers stay fixed during each GPU checkpoint.
• Every GPU can spike at once. This is a conservative budget; the four checkpoint values are not a simulated physical pulse or a proof over all rolling time windows.
• The selected PSU has the entered continuous rating at its actual temperature and input voltage. Verify 12 V rail capacity, connector availability, OCP allocation and manufacturer excursion testing.
• Certification-point interpolation and the default 75% efficiency at 5% load are estimates. A measured constant efficiency applies only where that measurement is representative.
• The 0.20 tariff is an illustrative user-editable price, not market data; cost uses the user's currency. Energy excludes off-state standby and external cooling.
• The default 0.80 circuit allowance is a planning input. All rigs can occupy the state with the larger modeled active/idle current simultaneously. The same entered true power factor applies in both states. Shared current is treated as continuous; breaker trip curves, inrush and local wiring rules are not modeled.
• One connector is diagnosed. Separate connectors share auxiliary watts equally, each of six supply paths and six return paths reaches a common node, and each branch has the same entered wire/other-path resistance. The default 6 mΩ series and 2 mΩ mating contacts are illustrative scenarios, not measured product data.
• Resistances and delivered voltage are fixed; PCB spreading resistance, temperature feedback, contact geometry and sense-pin behavior are excluded. The Kirchhoff/Joule equations are derived here, not an empirical failure correlation.

Use for single-phase PC and non-redundant workstation supplies within 90–264 V AC and the entered component/cable ratings. The modeled auxiliary connector is 12V-2x6; the current map must not be applied to legacy 6-pin/8-pin connectors or proprietary PSU-side pinouts. P/G labels identify modeled branches, not an official connector orientation or wiring diagram.

Intel's guide explicitly requires system-specific peak assessment for multiple cards. This tool screens four durations (100 µs through 100 ms); it does not certify ATX/PCIe compliance, model rail control loops or predict protective shutdown. The Intel envelope mode is intended for cards above 75 W. For a lower-power card, enter measured ratios. Efficiency beyond the selected PSU rating is only a held-endpoint extrapolation and must not justify overload operation.

Delivered connector voltage is restricted to 11.04–12.60 V using Intel Table 3-5 (p.28): +5%/−8% about 12 V. The 9.2 A default contact screen comes from Intel p.58 under its specified all-contacts-energized temperature-rise test. It is not a universal melting limit. Electrical I²R heat alone cannot give a contact temperature without thermal paths and material data. Verify actual assembly ratings, fully seated connections and manufacturer instructions. A voltage-drop warning means the fixed delivered-voltage scenario would require a PSU source above the modeled 12.6 V ceiling.

The circuit comparison is a preliminary load screen, not a breaker or wire-sizing certificate. Schneider Electric's 2014 discussion explains why continuous-load allowances depend on the listed installation. No single edition of electrical installation rules is automatically enforced. Check against the standard in force; installation work belongs to a qualified person. Source/profile and illustrative-price review date: 2026-10-25.