I was wondering about this a couple of days ago and decided to put it to the test with my M1 + mactop. Inference with InternVL 3.5 8b takes about 10W for 30-50 seconds on an average reply = 0.13Wh per inference, on 5yo hardware, but with a well-performing model.
I've done 2 of those in the past hour, while I also had 3 LED lamps on and used the microwave.
Inference: let's call it 2m at 10W = 0.33Wh
Lamps: 1h at 5W+7W+5W = 17Wh
Microwave: 1m at 900W = 15 Wh.
Total: 32.33 Wh
1% to gain knowledge
46% for food prep
53% for lights
mactop. Inference with InternVL 3.5 8b takes about 10W for 30-50 seconds on an average reply = 0.13Wh per inference, on 5yo hardware, but with a well-performing model.