“Historically, NVIDIA GPUs up to the B100 and B200 generations doubled their computational throughput (flop count) each time the numerical precision was halved.”