Press the button. Nothing happens. No fan spin, no board light. A brand new 700W power supply, wired exactly the way a model told me to wire it, sitting in the case like a brick with a fan screwed to it. That was the end of a two week hunt for a second RTX 3090, and it cost me a weekend.
The hunt
Used 3090s cost money, so I chased one on the marketplaces for a while. Two sellers flaked before I got anything, one promised a video and then disappeared, another sold the PC out from under me the same afternoon I said I'd take it. I passed on a $1,400 firm listing out of stubbornness and regretted it for a week. What landed was a Gigabyte AORUS 3090 picked up in person, cash in hand.
The first card is a 3090 Founders Edition, already in the case. The machine is a 5800X with 64GB of DDR4 in a Lian Li box, eight Corsair fans and a glass side panel, because two graphics cards behind glass are worth looking at. FE up top, AORUS bolted into the lower slot. One card for rendering, one for the models that judge the render, and I wanted them running at once, because a single 3090 doing both is a lesson in patience.
The plan I got handed
There was an 850W unit in the box. Two 3090s at 350W TGP each, plus a 5800X, is already over what I want to run on 850W, so I asked ChatGPT for the right shape and got back a confident, tidy answer. Buy a second power supply, run the Founders Edition off it, PCIe cables only, you don't even need the motherboard connections. Two cards, two supplies, load split. It read like something a person who builds machines would say.
So I bought a 700W unit and wired it exactly like that. The FE pulled its feed from the new supply, the rest of the machine stayed on the old one, I pressed the button, and there was no twitch anywhere on the desk.
What could never start
An ATX power supply does not turn itself on. It waits for PS_ON#, the green wire in the 24-pin connector, to be pulled low by the motherboard before it enables its rails. The board is what says go, and in that plan nothing was connected to the second supply but PCIe cables. No PS_ON link, no shared ground reference back to the first unit, so the rails never came up and the card stayed cold.
Every working dual-supply setup I've read about since uses a jumper on the second unit or an add2PSU style relay board that repeats the first supply's signal, and they tie the grounds together so both units share one reference. None of that was in the answer I got, because the answer came from something that has never had to make a machine boot.
The sizing problem underneath was worse. Ampere cards don't respect their own rating through load transitions. A 3090 sitting around 340W average pulls well above that in millisecond bursts, and two cards spiking at once is what a supply's overcurrent protection exists to catch. That's how renders die mid-run, or how a supply decides it's had enough. One unit with real headroom on a single rail is the sane shape, so the 700W went back to the store and a 1200W went on order.
One connector worth getting right. The Founders Edition uses NVIDIA's 12-pin MicroFit with an adapter to two 8-pin feeds, the 12VHPWR connector everyone argues about belongs to the 40-series cards, and the two get mixed up constantly in threads about power.
The guard that replaced the advice
The model didn't lie to me. It produced a plausible configuration, and I treated plausibility as verification because I wanted the build finished. The failure was mine, and the cost was a 700W paperweight and a weekend of my own time.
What I do now is cheap. Any hardware advice from a model gets converted into mechanism questions before a dollar goes out. What makes this start, what pulls which signal, what happens on the first load spike, what the thing does when one part is missing. If the answer can't survive those questions, the advice is decoration and I don't buy for it. For this build the check was one line. The supply needs PS_ON pulled low and a ground shared with the rest of the machine, and a second supply with neither is a box.
What shipped
The 1200W landed on the evening of 10/17. The box went down around 1:30pm that day for the swap, with the AORUS already seated in the lower slot, and I wouldn't be near the case when it came back, so I armed a watcher on the Mac. It polled ssh every 60 seconds, gave the machine 90 seconds to settle, restored the inference services, waited 45 more, then probed VRAM to find the card without the resident engine and relaunched ComfyUI on that one. A 4 minute health-check on the render server, a smoke render from the Mac, one message with the result. The watcher fired clean at 22:09, which means the machine booted about eight and a half hours after it went down and I never touched it.
Then the payoff I'd been building toward. On one card the render engine and the always-on inference engine were fighting for the same 24GB. A 1920x1088 image-to-video pass OOM'd, that path runs two forward passes per step, and 1664x960 at 50 steps fit only with the card clear, 21.6GB in use at 100% during sampling, at a two to three hour ETA. Worse, the resident engine kept creeping back onto the card mid-day, restarted by something following stale notes, and my first attempt to stop it used pkill -f over ssh, which matched the ssh session's own command line and killed my connection instead of the service. systemctl stop was the fix and that's now the rule on that box.
With two cards that class of problem goes away. The AORUS holds the resident engine at 21.4GB plus a 1.3GB vision sidecar, running 24/7, the other card holds the TTS model at 19.4GB with ComfyUI and the vision critic co-resident beside it, and a full render-and-critique round runs about 40 seconds. The grind that ran on top of it was 46 rounds across 6 briefs against a bar of 9, clearing twice at 8 out of 10 and never once hitting the bar, and the constraint left there is the models' taste.
Two 3090s on one 1200W supply, one switch, one signal, and no relay board to forget. The 700W went back to the store the morning I said it would, and the box has been running dual-card ever since.
