Sizing the Power Supply for a Multi-GPU AI Workstation

The power supply is the least glamorous part of an AI workstation and the one most often under-specified. An undersized or low-quality PSU does not just risk shutdowns — it can quietly destabilize a machine that otherwise looks perfect on paper, producing random reboots in the middle of long training runs that get misdiagnosed as software bugs for weeks. This guide explains how to size and select a PSU for a multi-GPU build that actually stays up.

Add up real draw, then leave headroom

Start from the rated board power of every component: each GPU, the CPU, drives, fans, and motherboard. A modern flagship GPU can draw 450 to 600W, and a Threadripper PRO around 350W, so a dual-GPU build is comfortably a four-figure wattage machine before you add the rest. Size the PSU to run at roughly 60 to 80 percent of its rating under full load. That keeps it in its most efficient, coolest, quietest band and leaves margin for the spikes below.

Transient spikes are the real sizing challenge

Modern GPUs draw brief power transients well above their rated TGP — spikes lasting only milliseconds but reaching far higher than the steady figure. Multiply that across two or four GPUs that can spike together and a PSU sized only to the nominal total will trip its protection and shut the system down under load. This is the single most common reason a multi-GPU workstation reboots randomly. Headroom is not waste here; it is what absorbs the transients.

Connectors: 12V-2x6, cables, and the melting-connector lesson

• Use native 12V-2x6 (the revised 12VHPWR) cables from an ATX 3.x PSU rather than daisy-chained adapters wherever possible.
• Seat each high-power connector fully — partial insertion caused the connector-overheating issues widely reported on early high-wattage cards.
• Verify the PSU has enough discrete GPU power cables for every card; do not split one cable across two hungry GPUs.
• For quad-GPU systems, confirm the chassis and PSU support the total connector count before committing to the layout.

Efficiency rating and why it pays back

An 80 PLUS Platinum or Titanium unit wastes less energy as heat than a Bronze one, which matters twice on a machine that runs near full load for hours. You pay less for power, and the PSU and the room both run cooler and quieter. On a 24/7 workstation the efficiency premium is recovered over the life of the machine, and the lower internal heat improves reliability of the supply itself.

Single-phase limits and circuit planning

There is a hard ceiling that catches people: a standard wall circuit can only deliver so much. A dense multi-GPU workstation can approach the limit of a typical office outlet, and several such machines on one circuit will trip a breaker. Plan the electrical environment alongside the build — dedicated circuits, the right outlet type, and UPS sizing — so the room can actually supply the hardware you are deploying.

Nexus Compute sizes every workstation PSU for true transient load, not just nominal draw, specifies native ATX 3.x power delivery, and validates the system under sustained stress before it ships — warranty-backed, quoted within 48 hours.