MATS @Anpaure breaks down how AI compute verification via side channels works: "When you have a GPU, it consumes different amounts of power depending on the workload. If it's inference or training, the wattage it pulls is very different." "If you have an external device like an oscilloscope and you hook it up to the GPU, you can draw a line, a graph of the power consumption over time, and it can be really granular. One of the tasks is looking at only that. You can probably tell apart…