3 ms·

We're not talking about live-training of models here, we're talking about executing a pre-trained model. CPU / power needs are MUCH lower.