[jorgemf]

One machine with 8 V100 GPUs. If you consider one TPU pod a single machine the TPU is faster. Those numbers also show that 8 GPUs are slower than 8 TPUs (so same conclusion as the article)