The outlet behind your desk is the weakest part of a two-GPU AI box. Two stock RTX 5090s and a CPU hold about 1,450 W, and your 1600 W supply pulls about 1,610 W from the wall to deliver that. A standard 15 A, 120 V circuit is good for 1,440 W of continuous load. Choosing the best PSU for a dual RTX 3090 or 5090 build starts at that outlet, then moves to the box.
We sized four builds from published numbers: the Intel ATX 3.1 power supply design guide, Cybenetics test reports, the spike captures from igorslab.de and TechPowerUp, and the maker spec sheets for every unit below. No PSU was load-tested in our lab, so every spike, efficiency and hold-up figure carries its source. You get the sizing math, the 12V-2×6 cable rules and the units rated for full output on a 120 V outlet, with listings and Cybenetics data pulled in late October 2026.
The PSUs compared
Every unit here is rated for its full output from 115 V to 240 V, and most drop to a lower rating somewhere between 100 V and 115 V, so a sagging outlet matters. The 115 V line rules out more 2000 W boxes than any other spec. The native 12V-2×6 count is the second filter, because two RTX 5090s need two of them.
| PSU | Rating, 115 to 240 V | Native 12V-2×6 | PCIe 8-pin | Cybenetics (115 V) | Price |
|---|---|---|---|---|---|
| MSI MEG Ai1600T | 1600 W | 2 | 9 on 7 cables | Titanium, A++ | about $620 |
| Seasonic PRIME PX-1600 ATX 3.1 | 1600 W | 2 | 6 | Titanium, A or A- (maker claim) | about $505 |
| Thermaltake Toughpower TF3 1650W | 1650 W | 2 | 9 | Titanium, Standard (loud) | about $450 |
| Super Flower Leadex Titanium 1700W | 1700 W | 2 | 6 | Titanium, A | about $410 |
| Corsair HX1500i (2025) | 1500 W | 0 (2 leads from 8-pin sockets) | 5 | Platinum, A- (maker) | about $320 |
| Corsair HX1200i (2025) | 1200 W | 0 (2 leads from 8-pin sockets) | 5 | Platinum, A- (maker) to A+ (review) | about $280 |
Read the Cybenetics column as the efficiency and noise rating at 115 V. Unmarked rows come from Cybenetics test reports, and rows marked maker are the claim on the spec sheet. Seasonic claims Titanium for the ATX 3.1 PX-1600 and lists both A and A- for noise in its own documents, but Cybenetics has no report for that revision yet. Prices are the Amazon US figures on our check date and they move weekly.
Step 1: Add up sustained watts, then the millisecond peaks
A GPU does not draw its board power as a flat line. It spikes for a few milliseconds at a time, and two cards can spike together. These are the published spike figures for the cards people actually pair for local LLMs:
| Card | Board power | Measured spike | Source |
|---|---|---|---|
| RTX 3090 FE | 350 W | 528 W for 1 ms at the stock limit | igorslab.de |
| RTX 3090 | 350 W | 485 W | TechPowerUp |
| RTX 4090 | 450 W | 508 W | TechPowerUp |
| RTX 5090 FE | 575 W | 644 W | TechPowerUp |
| RTX 5090 FE | 575 W | 824 W at 1 to 5 ms, 901 W under 1 ms | igorslab.de |
Note the gap between the two RTX 5090 rows. TechPowerUp samples too slowly to catch the shortest spikes, while Igor Wallossek measures at the millisecond level. Use the faster capture when you size, since your PSU sees every spike whether a slow meter records it or not.
Now add your CPU and the rest of the system. We assume 200 W for the CPU package and 100 W for the board, RAM, drives and fans, so 300 W outside the GPUs. That gives these four builds, with wall current worked out at about 90% efficiency, which is what Cybenetics measured for the 1600 W class near full load:
| Build | Sustained DC | Short peak | PSU | Wall current at 120 V |
|---|---|---|---|---|
| 2x RTX 3090, stock | 1,000 W | about 1,360 W for 1 ms | 1200 W floor, 1500 W comfortable | about 9.3 A |
| 2x RTX 3090 at 250 W | 800 W | about 1,230 W (estimate) | 1000 to 1200 W | about 7.4 A |
| 2x RTX 5090, stock | 1,450 W | about 1,950 W for a few ms | 1600 W with two 12V-2×6 | about 13.4 A |
| 2x RTX 5090 at 400 W | 1,100 W | about 1,700 W (estimate) | 1600 W advised | about 10.2 A |
Treat the capped peak figures as our estimates. We found no published spike captures for a power-capped 3090 or 5090, so we scaled the stock spikes by what Igor Wallossek measured on a 4090 with a software power limit on Windows, in proportion to each cap. Our dual RTX 3090 build cites Igor’s Lab advice of 460 to 500 W of capacity per card when a 3090 is overclocked to its 400 W limit, and the local AI workstation guide also lands on 1500 W for two 3090s plus a CPU.
One line in the Intel design guide changes how you read a box that says “5090 ready”. The ATX 3.x excursion budgets “assume only one PCIe Add-in Card will use these power excursions”, and the guide says it does not define a standard for systems with more than one. Your two-card AI box is outside what the sticker promises, so you size it yourself with the table above and leave headroom for both cards spiking at once.
Step 2: Check the wall circuit before the PSU
Electrical code limits a continuous load to 80% of the breaker rating. A PC running inference for hours is a continuous load, so in North America the numbers are:
- 15 A at 120 V: 1,440 W continuous (12 A)
- 20 A at 120 V: 1,920 W continuous (16 A)
- 230 V circuits: roughly double the headroom at the same current
Cybenetics ran full-load tests near 115 V on the 1600 W class, and the ASUS ROG Thor 1600W Titanium III drew 1,785 W from the wall to deliver 1,599 W, which is about 15.6 A. The Thermaltake TF3 drew 1,829 W, about 16 A. The code’s receptacle table caps a 15 A receptacle at 12 A of cord-and-plug load even on a 20 A circuit, so two stock 5090s at about 13.4 A need a 20 A circuit with a 20 A (NEMA 5-20) receptacle, ideally dedicated, or 230 V. The capped pair at about 10 A fits an ordinary outlet with nothing else on it.
Your next trap is the rating in the listing title. Several big units only hit their headline number on 230 V, and you have to read the 115 V line on the maker spec sheet to see it:
- SilverStone HELA 2050R: 1650 W at 115 V
- SilverStone HELA 3000Rz: will not run on 110/120 V
- Super Flower Leadex VIII Platinum 1600W: rated at 200 to 240 V only
- Cooler Master X Mighty Platinum 2000: 1600 W below 200 V
Do not trust the math alone either. Your own P3 Kill A Watt meter at about $30 shows you the real draw of the whole box under a long inference run. It is rated to 15 A, so it reads a capped build comfortably but tops out below the full-load draw of a 1600 W unit. If your box sits on a UPS, size that UPS for the wall figure, not the PSU label. Our homelab UPS guide tops out around 1,350 W, which suits the capped builds. A stock dual-5090 box needs a UPS rated about 2,000 W of real power on a 20 A input, such as an APC SMT2200 at 1,920 to 1,980 W, so check the watt rating rather than the VA figure.
Step 3: Pick a 1600 W unit for two RTX 5090s
Two stock 5090s need 1600 W, two native 12V-2×6 sockets and an ATX 3.1 design. A 1500 W unit runs the pair at 97% sustained load, with little margin for two cards spiking together. These four meet all three requirements.
MSI MEG Ai1600T PCIE5
If you want the best-documented 1600 W unit for two 5090s, buy this one. It is our first pick.
Cybenetics rates it Titanium and A++ at 115 V, the quietest result in this group, and its report shows a pass on the ATX 3.1 excursion test with a 21.3 ms hold-up time. MSI rates it 1600 W from 115 V to 240 V and 1200 W at 110 to 115 V, so check what your outlet actually measures with a meter. It has two native 12V-2×6 sockets and a 135 mm fan that stops below 55% load.
Its nine PCIe 8-pin plugs ride on seven cables, five with one plug and two with two, according to the Cybenetics cable list. That gives you six or more separate cable runs if you also run 8-pin cards, and the panel has room for those cables and both CPU EPS leads at once. It sold for about $620 from Amazon with only a few units left on our check date.
Seasonic PRIME PX-1600 ATX 3.1
Seasonic gives you a cheaper route to 1600 W with the PX-1600, and it carries a 12-year warranty.
Seasonic rates it 1600 W from 115 V to 240 V and 1300 W at 100 V, with two native 12V-2×6 sockets and six PCIe 8-pin. The weak point is the evidence: Cybenetics tested the older ATX 3.0 PX-1600 at 91.49% efficiency, but not this ATX 3.1 revision, so its Titanium efficiency and transient behavior are maker claims for now. Make sure the listing says ATX 3.1. The older ATX 3.0 model is still sold by third parties.
Expect to pay about $505 to $515 on Amazon. If noise matters more than money, about $650 buys the PRIME TX-1600 Noctua Edition, which tested Titanium and A++. The plain one, about $560, is the PRIME TX-1600.
Thermaltake Toughpower TF3 1650W
For the most connectors per dollar, you want the TF3, but it is loud under load.
You get two native 12V-2×6 sockets and nine PCIe 8-pin, so a mixed build with one 5090 and two 8-pin cards still has spare leads. Cybenetics passed it on the ATX 3.1 excursion test with a 23.5 ms hold-up, but rated its noise Standard at about 41 dBA, which is a lot for a desk-side box. It suits a basement rack or a closet better than a home office. Cybenetics listed only one 12+4 pin cable for its test sample while Thermaltake says two ship, so confirm both are in the box before you plan on two 5090s.
Thermaltake rates it in the manual at 1650 W from 115 V to 240 V and 1250 W at 100 to 115 V, and Cybenetics ran it at full power near 115 V. At 165 mm long it is also the shortest unit here. It sold for about $450 from Amazon.
Super Flower Leadex Titanium 1700W
Pick this one if you want headroom past 1600 W on a 115 V supply.
Super Flower rates it 1700 W from 115 V to 240 V, with two native 12V-2×6 sockets and six PCIe 8-pin. Cybenetics rated its Leadex ATX3+ 1700W Titanium and A, and its 12 V rail held 11.74 V at the 200% transient step. Its input label says 16 A at 115 V, and at a full 1700 W it would pull roughly 1,890 W from the wall. Every 1600 W unit here crosses the 15 A line at full load too, but this one goes furthest past it, so plan for a 20 A circuit with a 20 A receptacle, or 230 V.
Super Flower publishes one render for the Leadex Titanium ATX 3.1 family, so the image above is that family shot. It sold for about $410 from Amazon, less than the MSI, with few reviews yet.
Two more units fit this step. The ASRock PG-1600G passed the ATX 3.1 excursion test with a 20 ms hold-up and ships two 12+4 pin cables (we could not confirm the PSU-side socket type), at about $200 from the ASRock USA store on Amazon, a third of the price of the MSI. The FSP MEGA TI 1650W tested Titanium and A++ at about $500, but we could not read its maker spec page, so check the cable list before you buy.
Step 4: Pick a 1500 W unit for two RTX 3090s
Two 3090s sit at 1,000 W sustained, so 1500 W covers them with room for the 1 ms spikes. Your deciding spec is the 8-pin count, not the wattage.
Corsair HX1500i (2025)
Our dual 3090 build uses this unit, and it remains the pick if your two cards have two 8-pin inputs each.
Corsair rates it 1500 W from 115 V to 240 V and 1200 W at 100 to 115 V. It ships five PCIe 8-pin cables, which covers two Founders Edition or two-connector partner cards with one spare. Its two 12V-2×6 leads each start from two Type 4 8-pin sockets on the PSU, so they share that socket pool rather than adding to it.
Five cables is the trap. A pair of three-connector cards such as the ASUS Strix or EVGA FTW3 needs six 8-pin cables, so if you own those cards, step up to the MSI, which ships six or more separate runs, or check the PX-1600 cable list before you buy. It sold for about $320 from Amazon, and the RTX 3090 vs 5090 comparison covers which card to pair it with.
Step 5: Cap the cards with nvidia-smi and keep the cap after reboot
A power cap is the cheapest way to fit a two-card box on a normal outlet. Puget Systems measured four 3090s keeping 93% of their training speed at 250 W instead of 350 W, while the wall draw fell from about 1,570 W to 1,220 W. Why pay for a 20 A circuit if a cap costs you a few percent?
We took the commands below from the nvidia-smi documentation and the boot pattern Puget Systems published. We did not run them on a test box for this guide. First read the minimum and maximum limits your card accepts, since you cannot set a cap outside them:
nvidia-smi -q -d POWERPick a value between the Min Power Limit and Max Power Limit it reports, then enable persistence mode and set the cap on every GPU. Both need root, and 400 is the RTX 5090 example; use 250 for a 3090, since 400 would raise a 3090 above its stock limit:
sudo nvidia-smi -pm 1
sudo nvidia-smi -pl 400Neither setting survives a reboot, according to NVIDIA. A systemd oneshot unit applies them at boot, so create one:
sudo vim /etc/systemd/system/nvidia-power-limit.serviceAdd the unit below, changing 400 to the cap you chose. If your two cards need different caps, give each one its own line, such as ExecStart=/usr/bin/nvidia-smi -i 0 -pl 400 and ExecStart=/usr/bin/nvidia-smi -i 1 -pl 450:
[Unit]
Description=Set NVIDIA GPU power limit
After=systemd-modules-load.service
ConditionPathExists=/usr/bin/nvidia-smi
[Service]
Type=oneshot
ExecStart=/usr/bin/nvidia-smi -pm 1
ExecStart=/usr/bin/nvidia-smi -pl 400
[Install]
WantedBy=multi-user.targetReload systemd and enable the unit so it runs on the next boot and right now:
sudo systemctl daemon-reload
sudo systemctl enable --now nvidia-power-limit.serviceConfirm the limit took on each card. The power.limit column should show your cap:
nvidia-smi --query-gpu=index,power.limit,power.draw --format=csvYour cap trims the average far more than the spikes. On a capped 4090, Igor Wallossek measured a 40% lower limit cutting peaks by only about 17%, so keep the PSU sized for spikes even after the cap. Owners report that many RTX 5090 cards will not go below a 400 W limit, so check the Min Power Limit line before you plan a lower cap. For two 3090s at 250 W, the Corsair HX1200i (2025) at about $280 is enough. Watch the real draw over time with our DCGM, Prometheus and Grafana setup.
Step 6: Wire one cable per GPU socket
NVIDIA traced its first wave of melted 12VHPWR connectors to plugs that were not fully inserted, so this step protects your cards. 12V-2×6 is the ATX 3.1 replacement for 12VHPWR, and only the header on the card and the PSU changed: the power pins are 0.25 mm longer and the sense pins 1.5 mm shorter. A plug that is not fully seated now reads as open on both sense pins, and in ATX 3.1 that means 0 W instead of a reduced power level.
Look at the panel above to see what a native socket looks like: two 12V-2×6 headers, each feeding one card through one cable rated for 600 W. These are the ATX 3.1 changes that matter for a GPU build:
- 12V-2×6 header replaces 12VHPWR
- Open/Open sense now means 0 W
- Hold-up time relaxed to 12 ms at full load
- PSU excursion test levels unchanged from ATX 3.0
Shorter sense pins fix a half-inserted plug, but they do not fix uneven current on a fully seated one. der8auer measured 20 A or more on single wires of an RTX 5090 cable. Push each plug in until it sits flush with no gap, and keep the first 30 mm or so of your cable straight, as Corsair advises. Do you need an adapter for an 8-pin PSU? Use an 8-pin to 12V-2×6 cable from your PSU maker or a vendor that names your PSU model, never a generic one. CableMod, for example, sells a 2x 8-pin to 12V-2×6 cable for Seasonic and ASUS units at about $25.
Modular cables are not interchangeable, even inside one brand. Corsair says its Type 3 and Type 4 cables are not compatible with Type 5 units, and Thermaltake found two of its own 24-pin cables with completely different pinouts. For the HX1500i (2025), the matching cable is the Corsair Type 4 2x 8-pin to 12V-2×6 at about $19, not the Type 5 version at the same price.
For units with a native socket, the Corsair ThermalProtect 12V-2×6 cable at about $24 adds over-temperature protection, and Corsair says it fits any native 12V-2×6 socket. Thermaltake voids the TF3 warranty with third-party cables and ASUS says to use only bundled cables, so owners of those units should stay with the cables in the box. On 8-pin cards, run one separate cable per GPU socket. NVIDIA asks for four independent 8-pin cables on the 5090 adapter, and Seasonic advises separate cables above 225 W.
Datacenter cards are a separate trap. Tesla P40, P100 and V100 boards take a CPU EPS 8-pin, not a PCIe 8-pin, as our used datacenter GPU guide explains. Slot power is the last piece, and the multi-GPU motherboard guide covers boards with auxiliary slot power connectors.
Why two PSUs is the last resort
Builders who outgrow one PSU often add a second with an Add2PSU relay board at about $18, which switches the second unit on from the first. It works, but we found no Corsair, Seasonic or ASUS guidance on grounding or backfeed between two units, so you are on your own if something goes wrong.
Slot power always comes from the PSU that runs the motherboard, so a card fed by a second PSU splits its power between two supplies unless it sits on a powered riser. Forum builders treat that split as the thing to avoid, though no PSU maker documents it. Puget Systems asks for the two outlets of its dual-PSU workstation to be on separate 15 A circuits or one 20 A circuit, which brings you back to the wall again.
One 1600 W ATX 3.1 unit on a 20 A circuit with a 20 A receptacle, or on 230 V, is cleaner than two PSUs on a relay, and a power cap gets most builds there on an ordinary outlet. If even that is not enough, one larger card is often simpler than two, as our RTX PRO 6000 vs RTX 5090 comparison shows for a single 600 W card.
