If the Science paper is to believe, the first-generation TPUs have similar computation power at least, similar to a consumer-level GPU. It’s interesting to note that while they used 4 TPUs for the AlphaZero setup that played the match, they used 5,000 TPUs to train it!