Skip to main content
AI Gear Watch

Power, cooling, and noise

Updated 2026-08-15

Almost every write-up about local AI hardware stops at memory and compute. Then the box arrives, and the actual problems turn out to be that the power supply trips, the card cooks itself, or the thing is too loud to sit next to.

These constraints are knowable in advance. Here is how to check them.

Read the power number, then read the other power number

Vendors publish more than one figure and they do not agree with each other.

Take the RTX 5090. Its spec table lists Total Graphics Power (W) 575 and Required System Power (W) 1000. But NVIDIA’s own Founders Edition setup section on the same page says 850W Minimum, adding that “A higher power rating may be required depending on your system configuration” and telling you to check with your PSU vendor. The 1000 W footnote assumes a specific CPU pairing and notes that “Power requirements can be different depending on system configuration.”

Those two numbers are 150 W apart on one page. Neither is wrong; they answer slightly different questions. Plan against the higher one.

Our catalog records the card figure, and only the card figure, for GPU entries. A 575 W card is not a 575 W machine. Add the CPU, the drives, the fans, and the losses in the supply itself.

Whole-system entries are a different measurement

For Mac Studio, the GMKtec EVO-X2, and the other complete machines we track, the recorded power is for the entire computer. Apple’s figure for the M3 Ultra Mac Studio is a measured maximum continuous draw of 270 W for the whole box, idling at 9 W.

That is the comparison people get wrong. A 355 W RX 7900 XTX and a 270 W Mac Studio are not 85 W apart in what they ask of your outlet. One of those numbers is a component and one is a finished machine. See our methodology for how we keep the two straight.

The connector question

For the 5090, NVIDIA documents two wiring options: 4x PCIe 8-pin cables (adapter in box) OR 1x 600 W PCIe Gen 5 cable, and in the setup section, three 8-pin cables into the included adapter or 1x 450W or greater PCIe Gen 5 power cable. It advises a “PCIe CEM 5.1 compliant PSU”. It also asks you to plan 1.4" (36mm) additional space for power cables in the case, which is the kind of detail that turns a working build into a returned one.

Count the cables your supply actually has, not the ones it could have.

The passive-cooling trap

This is the single most expensive mistake in the used-datacenter-card market, and the vendor documentation states it plainly.

The H100 PCIe product brief lists Thermal solution: Passive, and explains what that means:

“It uses a passive heat sink for cooling, which requires system airflow to operate the card properly within its thermal limits.” (NVIDIA H100 PCIe GPU Product Brief)

A passive card has no fan. It is a heatsink and nothing else. It is designed for a server chassis whose own fans push a wall of air through it. Put it in a desktop case with gentle intake fans and it does not cool; it throttles, and then it protects itself.

The same is true of the A100, and it is true of any card whose spec sheet says passive. Tenstorrent sells its Blackhole in both flavors, active and passive, at the same price and the same specs; we track the active p150a deliberately, because for a desktop that is the one you want.

If you are buying a used enterprise card for a tower, budget for a shroud and fan kit, and accept up front that the result will be loud. This is not optional equipment.

The power modes nobody mentions

The H100 brief documents two distinct power configurations for the same card:

Total board power PCIe 16-pin 450 W or 600 W power mode: 350 W default, 350 W maximum, 200 W minimum. PCIe 16-pin 300 W power mode: 310 W default, 310 W maximum, 200 W minimum.

So the “350 W” card is also a 310 W card, and its floor is 200 W in either mode. That matters for two reasons: what you plan your supply around, and what you can do if the machine is thermally cornered. Many cards can be power-capped well below their rating, and the throughput cost is often smaller than the wattage cut. That is worth measuring on your own hardware.

We record the higher-mode figure in the catalog, since that is the number to size a power supply against.

Sizing the room, not just the machine

Watts do not disappear. Essentially all of the power a machine draws leaves it as heat, into the room you are sitting in.

Apple publishes this directly for Mac Studio: alongside Max 270 W it lists 921 BTU/hr. That is the only vendor in our catalog that states the thermal output explicitly, which is a shame, because it is the figure that tells you whether a small office becomes uncomfortable.

The rough shape of the problem: a machine pulling several hundred watts continuously is a space heater running continuously. If it lives in a closet, a small room, or anywhere without air exchange, that is a real constraint on where the hardware can go. Plan the location before the purchase.

What we are not going to tell you

No decibel figures. None of the vendors in our catalog publish noise ratings for these products, and the numbers circulating in reviews are measured at distances and in rooms that are not yours. We would be passing along someone else’s measurement as if it described your desk.

No idle-power numbers except Apple’s, which is the only one published (Idle 9 W for the M3 Ultra Mac Studio). Idle draw varies enormously with driver version, connected displays, and what the machine is doing when you think it is doing nothing.

No efficiency curves, no perf-per-watt rankings. Those require throughput measurements, and we do not publish throughput numbers we did not measure. See what can you actually run at home for what we do publish and why.

Measure your own machine. A cheap wall meter tells you more about your setup in an afternoon than any spec sheet will.

The short version

Before you buy: check whether the card is passive, count your PSU’s cables, plan against the vendor’s higher power figure rather than its lower one, and decide where the heat is going to go. All four are answerable from documentation before any money moves.

Compare the power figures across everything we track on the hardware list, keeping in mind which entries are cards and which are complete machines.

All guides