3 ms·

I'd probably consider the individual inference machine (GPU, TPU, etc.) to be the deodand in that case.