Google has officially introduced the second generation of their specialized Tensor Processing Unit. The new TPUs are head and shoulders above the old ones, and Google’s way of implementing them for the Google Cloud Platform, being a stack of 64 of them daisy chained for cooperation, provides up to 11.5 petaflops of raw computing muscle. The raw power behind these new TPUs simply cannot be understated. Google gave a fairly good example in their blog post announcing the new units, saying that one eighth of a TPU Pod, or just 8 of the new TPUs working in concert, managed to complete a machine learning task in one afternoon that used to take a full day for an array of 32 of the top GPUs on the market. Integration in Google Cloud Platform means that clients can build machine learning applications on this incredibly powerful hardware with no upfront investment.
The new TPUs are purpose built for machine learning tasks, but Google has made it easier than ever for those who want to use them on top of the Google Cloud Platform to program them using TensorFlow. These Cloud TPUs are integrated with off-the-shelf hardware like Intel CPUs and NVIDIA GPUs for compatibility with a client’s hardware stack, and bring all of that incredible power straight to whatever a client is working on. Naturally, the new TPUs are going to eventually be implemented in many of the places that Google is currently using first generation models, as well.
Google topped the whole thing off by announcing the TensorFlow Research Cloud. It’s an array of 1,000 2nd generation TPUs working in concert that’s shared among qualifying machine learning researchers. This means that there’s a whopping 180 petaflops of computational power to go around, for qualifying researchers. Each Cloud TPU puts out 180 teraflops, so even one can utterly decimate most off-the-shelf hardware that researchers may be able to snag out of their own pocket. The kicker is that qualifying researchers will get in on all of this power completely free of charge, and can build whatever they want on it. Google’s latest humanitarian effort to boost tech innovation isn’t available yet, but researchers who are interested can already sign up, and will be alerted when the system is available for them to use. One thing that was left a bit unclear is exactly what qualifies a researcher or group of researchers for access to the TensorFlow Research Cloud; most likely, this will wind up being a discretion call for Google staff, based on how much they think a given researcher or team could benefit from a share of the TensorFlow Research Cloud.