Yes, a UPS can assist by providing backup power during outages, which may prevent crashes caused by power loss.
Yes, a UPS can assist by providing backup power during outages, which may prevent crashes caused by power loss.
I have two 4090s operating from a 1650 watt power source. The system frequently stops but only when starting up tasks on the GPUs. It seems power increases as computation begins, triggering crashes. Usually after the job starts it completes without issues, though it often fails right at startup. I suspect temporary voltage spikes could be the cause. A UPS might help, but I'm not sure. Removing one GPU didn't resolve the problem, so replacing it with a cheaper unit isn't ideal. You mentioned crashes are worse in colder weather. Today on a hot day it fails every time you try to launch a large model on the GPU. Your build is under stress.
Are the GPUs running hot or is it the CPU that's causing issues?
The computer fails even during its initial use, especially when both the GPU and CPU run at low temperatures and minimal power. It seems overheating isn't the main issue—the system doesn't have enough time to overheat before crashing. I also ran memory tests, which helped shape my current idea. My hypothesis is that the power supply cuts off during a voltage drop to safeguard the device, which is why I considered using a UPS. I brought up the heat topic because hot conditions from air conditioners and similar devices can strain the power grid.
I recall there was definitely something related to temporary voltage spikes and the 4090 along with several PSUs. It’s highly probable your PSU is the culprit, and a UPS won’t help much. In reality, most UPS systems are designed to do nothing by default. You’ll notice this until you invest a significant amount in a UPS; then it will act like an expensive power strip. Under normal conditions, it behaves just like any regular power strip, switching seamlessly within 10ms from the wall outlet to an internal inverter when power is lost. If you’re spending a lot on a UPS, you should opt for one that converts all incoming power directly to DC before sending it through batteries and then to a massive inverter—this continuous conversion prevents any switching. Such online upgrades are ideal if you suspect wall power issues or voltage sag.
If your PSU is the Thermaltake ToughPower GF3 1650W, it’s an ATX3.1 model, which should handle transient spikes fine. It also has an OPP rating of 2000W, so it can easily absorb two 4090 spikes at roughly double their draw for a brief moment. What about your CPU load? Is this for personal use or professional work? Honestly, my advice is to go with a bigger PSU in the 1800-2400W range. The 4090 is definitely powerful and shouldn’t be underestimated.
240 volts seems a bit unusual for a home setup—120 volts is standard for most shared circuits. It’s worth considering whether upgrading to 240 volts would actually save money compared to a high-end UPS, especially since you’re already on a 20 AMP circuit. For your machine learning work and hobby tasks, a reliable PSU that matches your needs is key. You might want to explore options that balance performance and cost without overcomplicating your setup.