Browsers Using AI (incremental power usage?)

[bdub76]

Has anybody measured the incremental power usage for browsers with AI enabled as discussed by Steve on Security Now?

From my experience, I expect AI training to cost far more than usage, but I still expect an incremental increase in power usage. What I don't know is how much more power is being used.

 

[PHolder]

Most CPUs and GPUs are not designed for a low power mode for only a portion of them... they tend to be something like "cut back power to all cores to half" or something like that. (I suspect it's actually a hard computer science problem to design a small (i.e. cost effective) silicon slab that can be selectively cooled properly if only a portion of it were working really hard.) So anyway, the question becomes whether the CPU/GPU/NPU is using more power because it would otherwise be idle, or if it's just going from say 25% utilization toward something higher (up to 100% obviously ). So maybe it will cost more, but maybe it's not significantly more... like 10% more. The question is, is even 10% more a good return to the user... for me, I would suggest not right now until AI is actually trustworthy.

 

[SeanBZA]

Modern multicore units will slow down idle cores, and also skew down the voltage to them dynamically. So an AI load, which will tend to hog all cores, and all threads, along with massive amounts of memory access, will definitely ramp the entire CPU unit up to full power, till it hits thermal limits and starts to throttle back.

 

[davidgrouse133]

Has anybody measured the incremental power usage for browsers with AI enabled as discussed by Steve on Security Now?

From my experience, I expect AI training to cost far more than usage, but I still expect an incremental increase in power usage. What I don't know is how much more power is being used.

Same here, Chrome’s AI junk adds ~1 W continuous idle draw even in background tabs. Edge Copilot worse (~2 W).
Kill the flags and it’s gone. Firefox still clean.