Solely every week after Nvidia’s new AI-focused Volta GPU structure was introduced, Google goals to steal a few of its thunder with its new, second-generation, Tensor Processing Unit (TPU) that it calls a Cloud TPU. Whereas its first technology chip was solely appropriate for inferencing, and due to this fact didn’t pose a lot of a risk to Nvidia’s dominance in machine studying, the brand new model is equally at residence with each the coaching and working of AI techniques.
A brand new efficiency chief amongst machine studying chips
At 180 teraflops, Google’s Cloud TPU packs extra punch, at the least by that one measure, than the Volta-powered Tesla V100 at 120 teraflops (trillion floating level operations per second). Nonetheless, till each chips can be found, it gained’t be potential to get a way of an actual world comparability. Very similar to Nvidia has built servers out of multiple V100s, Google has additionally constructed TPU Pods that mix a number of TPUs to realize 11.5 petaflops (11,500 teraflops) of efficiency.
For Google, this efficiency is already paying off. As one instance, a Google mannequin that required a whole day to coach on a cluster of 32 high-end GPUs (in all probability Pascal), will be skilled in a day on one-eighth of a TPU Pod (a full pod is 64 TPUs, so which means on eight TPUs). In fact, commonplace GPUs can be utilized for all kinds of different issues, whereas the Google TPUs are restricted to the coaching and working of fashions written utilizing Google’s instruments.
You’ll be capable of lease Google Cloud TPUs to your TensorFlow functions
Google is making its Cloud TPUs obtainable as a part of its Google Compute providing, and says that they are going to be priced much like GPUs. That isn’t sufficient data to say how they may evaluate in price to renting time on an Nvidia V100, however I’d anticipate it to be very aggressive. One downside, although, is that the Google TPUs at the moment solely assist TensorFlow and Google’s instruments. As highly effective as they’re, many builders won’t wish to get locked into Google’s machine studying framework.
Nvidia isn’t the one firm that must be nervous
Whereas Google is making its Cloud TPU obtainable as a part of its Google Compute cloud, it hasn’t mentioned something about making it obtainable exterior Google’s personal server farms. So it isn’t competing with on-premise GPUs, and definitely gained’t be obtainable on aggressive clouds from Microsoft and Amazon. Actually, it’s prone to deepen their partnerships with Nvidia.
The opposite firm that ought to in all probability be nervous is Intel. It has been woefully behind in GPUs, which suggests it hasn’t made a lot of a dent within the quickly rising marketplace for GPGPU (Normal Objective computing on GPUs), of which machine learning is a big half. This is only one extra manner that chip that would have gone to Intel, gained’t.
Huge image, extra machine studying functions can be transferring to the cloud. In some circumstances — for those who can tolerate being pre-empted — it’s already cheaper to lease GPU clusters within the cloud than it’s to energy them regionally. That equation is simply going to get extra lopsided with chips just like the Volta and the brand new Google TPU being added to cloud servers. Google is aware of that key to rising its share of that market is having extra vanguard software program working on its chips, so it’s making 1,000 Cloud TPUs obtainable totally free to researchers prepared to share the outcomes of their work.