Yeah, it is a lot, but it is the same in "fully working" cars.
Additionally, once you to the math things look bad: 8% of a battery pack of 60kWh is around 5kWH. 5000 Wh / 24h = 208 W.
200W to keep few cameras on and a small CPU/GPU to run the logic seems a lot. Especially considering that you don't need to analyze every frame of every camera. In fact, you could easily run one frame every 5 seconds making it ~ 1FPS analyzed (6 cameras) since the car is parked.
Additionally, you don't need to run heavy models on all the images. You can run some sort of diff to see if pixels changed within a tolerance threshold between the new and the old image and only if that's true then you run the ML models.
From my understanding the problem is that the computer just....doesn't sleep if you have Sentry mode running. It's not that it's sleeping for a few seconds and then waking up, it's just fully awake drawing it's normal base load, which for HW4 [1] appears to be ~300W.
A good way to verify this if you've got a Tesla is how long it takes for the car to acknowledge commands sent via the mobile app. If they're near-instant (ignoring TCP latency ofc), the car is "Awake" and drawing it's full load. If they take 5-15 seconds, the car is Asleep and only polling it's LTE antennas for push notifications every 10s or so. If they timeout entirely, the car is in Deep Sleep and drawing almost no wattage (at least based on my observations with my wife's M3).
When in Sentry mode, responses are ALWAYS instant, so the car is fully Awake and drawing full wattage.
Thanks for point that out, but I then believe it's a combination of the two.
I don't believe that an embedded PC that allegedly has the same power as a PS5 consumes 200/300W idling. Even if it doesn't sleep it should consume in the order of 10s of W, not 100s. For reference, my PC here has a 3070, a ryzen 7 7xxx (I don't remember the exact model) and consumes like 40W at the outlet when idling.
So I am willing to assume that yes, the onboard PC doesn't idle when sentry mode is active, but the 200+ W draw is due to it running terrible software, and not just not entering CPU states that are energy efficient.
Additionally, once you to the math things look bad: 8% of a battery pack of 60kWh is around 5kWH. 5000 Wh / 24h = 208 W.
200W to keep few cameras on and a small CPU/GPU to run the logic seems a lot. Especially considering that you don't need to analyze every frame of every camera. In fact, you could easily run one frame every 5 seconds making it ~ 1FPS analyzed (6 cameras) since the car is parked.
Additionally, you don't need to run heavy models on all the images. You can run some sort of diff to see if pixels changed within a tolerance threshold between the new and the old image and only if that's true then you run the ML models.