A Reminder, Guys, Undervolt your GPUs Immediately. You will Significantly Decrease Wattage without Hitting Performance.
Undervolting an RTX 3090 and 5070 Ti slashed power draw by 80-100W while maintaining full performance.
A viral Reddit post by user Iory1998 is serving as a powerful reminder for AI developers and PC enthusiasts about the significant benefits of GPU undervolting. Using the popular tool MSI Afterburner, the user demonstrated how to manually lower the voltage supplied to their graphics cards, resulting in dramatic efficiency gains. Their specific setup, featuring a water-cooled NVIDIA RTX 3090 and an air-cooled RTX 5070 Ti, saw power consumption drop by roughly 80-100W per card under full load, with the 3090 drawing 290-300W instead of 350-380W and the 5070 Ti using 180-200W instead of 250-300W.
Crucially, this substantial reduction in power draw and subsequent heat output came with no penalty to computational performance. The user reported stable operation and even a slight increase in frames per second (FPS) during gaming on the 5070 Ti. This efficiency translated into much lower operating temperatures, with the tightly sandwiched cards staying below 60°C and 50°C respectively, aided by custom fan curves set via FanControl software. The technique highlights an often-overlooked method for users running intensive AI model training, rendering, or gaming to reduce electricity costs, system noise, and thermal stress on hardware.
The process involves carefully lowering the voltage in small increments while stress-testing the GPU for stability, a practice that can extend the lifespan of expensive hardware and make high-performance computing more sustainable. For professionals running multi-GPU servers or local AI inference setups, these savings can compound significantly, reducing operational costs and cooling requirements without sacrificing the speed needed for tasks like running Llama 3 or Stable Diffusion models.
- Undervolting an RTX 3090 cut its power draw from 350-380W to 290-300W, a reduction of ~80W (23%).
- An RTX 5070 Ti's consumption dropped from 250-300W to 180-200W, saving ~70W (28%) with no performance loss.
- Temperatures for the tightly packed GPUs stayed under 60°C, reducing thermal stress and potential cooling costs.
Why It Matters
For AI professionals running local models, this technique slashes electricity costs and cooling demands for GPU clusters without impacting compute speed.