PSU Sizing Guide for LLM Rigs: How Many Watts Do You Actually Need?

TL;DR: Single RTX 4090 build needs 850W minimum, 1000W recommended. Dual RTX 3090 build needs 1200W minimum, 1600W recommended. For LLM rigs specifically, buy more headroom than a gaming system because inference loads are sustained — not bursty like gaming. An 80 PLUS Gold or Platinum PSU is non-negotiable for 24/7 operation. The Seasonic Focus GX-1000 and Corsair RMe RM1000e are the picks under $200.

Why LLM Rigs Are Harder on PSUs Than Gaming Rigs

A gaming system spikes power during intense scenes, but the average draw is much lower. A local LLM inference server running continuous requests sustains near-peak GPU power draw for hours at a time. That's a fundamentally different stress profile.

A cheap PSU that handles 15-minute gaming sessions might fail or throttle after 4 hours of sustained 400W inference. Worse, PSU efficiency drops sharply under light loads — a 1000W PSU running at 30% load (300W) on a cheap 80 PLUS Bronze unit is measurably less efficient and runs hotter than the same workload on a Gold-rated unit. For a machine you leave on all day, that heat and efficiency difference costs money.

Build for the actual workload: sustained, high-load, often overnight.

Step 1: Calculate Your Peak Power Draw

Add up the TDP (Thermal Design Power) of your components:

GPUs are the dominant draw:

• RTX 5090: ~575W TDP
• RTX 4090: ~450W TDP
• RTX 3090: ~350W TDP
• RTX 4080 Super: ~320W TDP
• RTX 4070 Ti Super: ~285W TDP
• RTX 3080: ~320W TDP

CPU:

• Ryzen 9 9950X: ~170W TDP (can spike to ~230W)
• Core Ultra 9 285K: ~125W TDP (can spike to ~250W depending on board power limits)
• Ryzen 5 7600: ~65W TDP
• Threadripper 9970X: ~350W TDP

Everything else (RAM, storage, fans, motherboard): Budget $50-80W total for a standard ATX build.

Add it up, then apply a headroom multiplier.

Step 2: Add Headroom

PSUs are most efficient at 50-80% load. Running a PSU at 95% load degrades its lifespan and can cause voltage ripple that destabilizes your system. For sustained LLM workloads, aim for your actual peak draw to land at 70-75% of PSU rated capacity.

Headroom formula: (Peak draw) ÷ 0.75 = minimum PSU rating

Single RTX 4090 build:

• GPU: 520W (real-world peak)
• CPU: 180W
• Rest: 75W
• Total peak: ~775W
• ÷ 0.75 = 1,033W → get a 1000W PSU minimum

Single RTX 3090 build:

• GPU: 390W
• CPU: 180W
• Rest: 75W
• Total peak: ~645W
• ÷ 0.75 = 860W → 850W works, 1000W is comfortable

Dual RTX 3090 build:

• GPUs: 780W combined
• CPU: 180W
• Rest: 100W
• Total peak: ~1,060W
• ÷ 0.75 = 1,413W → 1600W is the safe choice, 1200W is minimum

Step 3: Choose the Efficiency Rating

80 PLUS Bronze — minimum acceptable. 82-85% efficiency at 50% load. Fine for gaming, marginal for 24/7 LLM inference.

80 PLUS Gold — the right choice for most LLM rigs. 88-92% efficiency. The price premium over Bronze pays for itself in electricity savings over 6-12 months of 24/7 operation.

80 PLUS Platinum/Titanium — diminishing returns for most builds. Makes sense for professional inference servers running years at a time, or if you're in a location with expensive electricity ($0.25+/kWh). The Seasonic Prime TX-1000 Titanium runs ~$279 — justified for a production server, overkill for a personal rig.

Specific PSU Recommendations

Budget to Mid-Range: Corsair RMe RM850e (850W) — ~$95

ATX 3.1 compliant, 80 PLUS Gold, fully modular. Good for a single mid-range GPU build (RTX 4070 Ti and below). Corsair's reliability reputation is well-earned. The RMe series uses Japanese capacitors and has a 7-year warranty — meaningful for a 24/7 machine.

Best All-Rounder: Seasonic Focus GX-1000 (1000W) — ~$149

80 PLUS Gold, fully modular, 10-year warranty. The Focus GX series has been a trusted choice in workstation builds for years. At 1000W, this covers a single RTX 4090 build with comfortable headroom, and stretches to cover some dual-GPU builds with lower-TDP GPUs.

Seasonic manufactures PSUs for other brands too (including some Corsair lines) — buying Seasonic direct means you know what you're getting.

Dual GPU Choice: Corsair HX1500i (1500W) — ~$279

For dual RTX 3090, dual RTX 4090, or any single-card build where you want massive headroom. 80 PLUS Platinum, fully modular, includes digital monitoring via iCUE. The 1500W rating handles nearly any consumer multi-GPU configuration.

The iCUE monitoring is actually useful for server builds — you can track PSU efficiency, temperatures, and power delivery over time without separate power monitoring hardware.

SFX Builds: Corsair SF1000 (1000W) — ~$199

For small-form-factor builds using smaller cases. The SF series is 80 PLUS Platinum and handles an RTX 4090 in a compact chassis. SFX PSUs run warmer than ATX units at sustained loads — make sure your case has airflow past the PSU bay.

ATX 3.1 and PCIe 5.1: Do You Need It?

Newer PSUs carry ATX 3.1 and PCIe 5.1 certification, which adds native 12V-2x6 (12VHPWR) connectors for RTX 40/50 series GPUs.

Older PSUs with adapters have caused connector damage on RTX 4090 cards — the adapter doesn't always seat perfectly, causing heat buildup and in rare cases melted connectors. If you're buying a new PSU for an RTX 4090 or 5090, get one with native 12V-2x6 connectors rather than using an adapter. ATX 3.1 certification is your signal that the native connector is included.

The "Just Buy More" Trap

There's a counterintuitive trap here: people see that PSUs are most efficient at 50-80% load and conclude they should buy as large a unit as possible. A 2000W PSU running at 30% load is actually less efficient than a 1000W PSU running at 60% load — efficiency at very low load percentages is poor on most units.

Buy a PSU that puts your typical workload in the 50-80% zone. If your single RTX 4090 rig draws 700W during inference, a 1000W PSU (70% load) is ideal. A 1600W PSU (44% load) is slightly less efficient and costs more for no benefit.

See Also

• Best GPUs for Local LLMs 2026
• Local LLM Power Consumption Cost Guide
• Multi-GPU LLM Inference: How to Split Large Models Across GPUs