PSU Calculator

Size your power supply for AI inference

Multi-GPU AI rigs have unusual power profiles: GPUs spike to full TDP during inference, and a 2× RTX 3090 build can draw 800–900W sustained. Undersizing a PSU is a fire risk; oversizing wastes money on idle efficiency. The full calculator will size PSU, circuit breaker, and UPS for your exact component list.

Full calculator in development

The interactive PSU sizing tool is coming. In the meantime, use the quick reference table below for GPU TDP values and the build planner for complete part lists including PSU.

Open the build planner — it includes PSU recommendations for each build profile.

GPU power reference (TDP at rated load)

TDP figures are manufacturer-rated. Real inference load often reaches 90–100% TDP. Size your PSU to: (GPU TDP + CPU TDP + drives + fans + 20% headroom). For 80+ Gold and above at rated efficiency, multiply system draw by 1.1 for AC wall draw.

GPU Tier TDP (W) Single GPU PSU min Dual GPU PSU min
RTX 4060 Ti 8GB / 16GB Budget 165 W 450 W 850 W
RTX 4070 Ti Super 16GB Mid 285 W 600 W 1150 W
RTX 4080 Super 16GB Mid 320 W 650 W 1200 W
AMD RX 7900 XTX 24GB Sweet spot 355 W 700 W 1300 W
RTX 3090 24GB Sweet spot 350 W 700 W 1300 W
RTX 4090 24GB Sweet spot 450 W 800 W 1500 W
NVIDIA A6000 48GB Workstation 300 W 600 W 1150 W
RTX 6000 Ada 48GB Workstation 300 W 600 W 1150 W

Single GPU PSU min = (GPU TDP + 200W system) × 1.2 headroom, rounded to next 50W. Dual GPU PSU min = (2× GPU TDP + 350W system) × 1.2, rounded to next 50W. CPU TDP assumed 125W; add more for Threadripper/EPYC. Data from manufacturer spec sheets — TDP is thermal design power, not guaranteed peak.

Quick sizing rules (until the calculator is live)

80+ Gold minimum for any AI rig
Gold (87–90% efficient) reduces AC draw, heat, and electricity cost meaningfully over Bronze. Platinum and Titanium are worth it for 24/7 inference servers.
20% headroom over peak draw
PSUs run most efficiently at 50–80% of rated load. A 1000W PSU running a 700W system is in the sweet spot. Under 50% or over 90% are both suboptimal.
Dedicated 20A circuit for 600W+ builds
A 20A household circuit supports ~1800W at 120V, but shared circuits have other loads. Any build drawing 600W+ under sustained inference load should be on a dedicated circuit.
UPS sizing: runtime × load + 10% margin
For a 900W inference server, a 1500VA/900W UPS gives ~5–8 minutes of runtime — enough for clean shutdown. For longer backup, size up or use a 2000VA+ unit.
RTX 3090 / 4090 power spikes
Ada/Ampere GPUs can spike above TDP for short intervals. NVIDIA recommends a PSU headroom of at least 20% above rated TDP. On a 450W 4090, plan for a minimum 550W GPU budget.
Build Planner

Build your AI rig parts list

The build planner includes PSU recommendations for each rig profile — use it while the PSU calculator is in development.

Open Calculator