Ben Gilbert: evaluation

Transcript context

…Right. Why would you tell anybody about this? Because it’s not like they’re offering these in Google Cloud, at least at first. Why would you want to tell the rest of the world what you’re doing? The whole thing was a complete secret for at least a year before they announced it at Google IO. So really crazy. The other thing to know about the TPUs is they were done in time for the AlphaGo match. That match ran on a single machine with four TPUs in Google Cloud. Once that worked, obviously that gave Google a little bit of extra confidence to go really, really real production. That’s the TPU. V1 by all accounts was not great. They’re on V7 or V8 now. It’s gotten much better. TPUs and GPUs look a lot more similar than they used to, than they’ve adopted features from each other. But today, Google it’s estimated has 2–3 million TPUs. For reference NVIDIA shipped—people don’t know for sure—somewhere around 4 million GPUs last year. People talk about AI chips like it’s this one horse race with NVIDIA. Google has an almost NVIDIA-scale internal thing making their own chips at this point for their own and for Google Cloud customers. The TPU is a giant deal in AI in a way that I think a lot of people don’t realize. This is one of the great ironies and maddening things to OpenAI and Elon Musk, is that OpenAI gets founded in 2015 with the goal of, hey, let’s shake all this talent out of Google and level the playing field, and Google just accelerates.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.