“Google's TPU-AT delivers nearly three times the compute performance per pod over previous generations due to native NLXU quantization.”