Our current computer vision training workstation, at full train, seems to pull 450-500 watts (per nvidia-smi and the rated TDP of our CPU).
At 120 training days, and 24 hours in a day, I believe that works out to ~1,400 kilowatt hours. However, I’m stupid with electrical stuff (and most stuff, honestly), so I’m happy to be corrected.
Our current computer vision system uses an EVGA SuperNOVA 1200 P2 power supply, which was the most efficient power supply I could get my hands on in 2017, back when cryptocurrency mining was taking over the world and efficient power supplies were very hard to source. It’s Platinum certified, which means it’s ~94% efficient at half-load (meaning we would need to draw 480 watts from the wall to fulfill 450 watts of need). The new computer vision system we’re building will have a titanium certified power supply, which means it’s at least ~96% efficient at half-load. So we’d only need to draw 470 watts from the wall to fulfill the same 450 watt need. Titanium is the highest certification rating for computer power supplies.