Running a local language model, generating batches of images or fine-tuning a model can keep a graphics card working flat out for hours. If you'd like it to draw less power and run cooler during those jobs, you can cap its power with nvidia-smi, the command-line tool NVIDIA installs with its driver. This guide follows NVIDIA's nvidia-smi documentation. Forge PC hasn't tested power limits on specific cards, so we can't tell you how much speed you'd lose at a given limit; test your own workload.
Why cap power for AI work?
NVIDIA's guide to LLM inference explains that the decode phase, where a model generates its reply one token at a time, is memory-bound: the speed at which data moves from memory dominates, rather than how fast the computation happens. That suggests a modest power cap may cost less speed in text generation than in compute-heavy work such as image generation or training, but it depends on the card and the job, so measure before and after. If you also game on the same PC, our GPU undervolting guide covers a different way to cut power with MSI Afterburner.
Step 1: Open an administrator terminal
NVIDIA's documentation says changing the power limit requires root, which on Windows means running as administrator. Right-click the Start button, choose Terminal (Admin) and accept the prompt.
Step 2: Check your card's allowed range
Type:
nvidia-smi -q -d POWER
The -d POWER option shows only power information. Look for these lines, which NVIDIA defines as:
Current Power Limit: the limit the power management algorithm is enforcing now.
Default Power Limit: the card's standard limit.
Min Power Limit and Max Power Limit: the lowest and highest values you're allowed to set.
Note the default figure, as you'll need it to undo the change. If you have more than one GPU, nvidia-smi lists each with an index number (0, 1 and so on).
Step 3: Set a lower limit
Use the -pl (power limit) option with a value in watts, and -i to choose the GPU. For example, to set GPU 0 to 250W:
nvidia-smi -i 0 -pl 250
NVIDIA says the value must be between the Min and Max Power Limit that nvidia-smi reports, and that power limits are only available on supported devices. If your card or driver doesn't support changing it, use the power limit slider in MSI Afterburner instead; our Afterburner overclocking guide shows where it is.
A sensible approach is to start a little below the default and work down in steps, running the same AI job each time and noting how long it takes.
Step 4: Watch what the card is doing
To check the limit is being applied while your job runs, repeat the power query every five seconds:
nvidia-smi -q -d POWER -l 5
Press Ctrl+C to stop. NVIDIA's documentation describes the readings as the average and instantaneous power draw for the whole board. If the card is hitting your cap, NVIDIA says the "SW Power Cap" throttle reason shows the driver reducing clocks because the GPU is consuming too much power. For temperatures and clocks in a friendlier view, use HWiNFO.
Step 5: Put it back
Set the limit back to the Default Power Limit you noted earlier. For example, if your card's default was 320W:
nvidia-smi -i 0 -pl 320
NVIDIA's documentation doesn't say whether a power limit set this way survives a restart (it notes that persistence mode, a separate setting, doesn't), so run nvidia-smi -q -d POWER after rebooting to see which limit is active and re-apply it if needed.
Things to keep in mind
A lower power limit reduces the maximum the card can draw; it doesn't replace a power supply that's too small. Our PSU guide covers sizing.
If you run two GPUs for local LLMs, set each one with its own -i index.
Don't set a limit above the card's default to try to overclock it: that's outside the scope of this guide, and the Max Power Limit is the most the driver allows.
Our verdict
Open Terminal as administrator, run nvidia-smi -q -d POWER to see your card's default, minimum and maximum limits, then set a lower cap with nvidia-smi -i 0 -pl <watts> and check it with the looping power query. Test your own AI workloads at each step, and note the default so you can put it back. Forge PC hasn't tested specific limits on specific cards.