In conjunction with this year's Google I/O 2018 Developers Conference, Google has now revealed its third generation of Tensor Processing Units (TPU) for use in its "TPU Pods." For those who may not already know, those Pods are Google's supercomputer arrays purpose-built for machine learning (ML) applications. Specifically, TPUs are utilized to train A.I. on an accelerated timeline. For comparison, the company's second-generation Pods were comprised of no fewer than 64 TPUs and provided up to 11.5-petaflops. That's already very impressive but the new TPU Pod is exponentially more powerful, offering more than 8 times the power with more than 100 petaflops of ML hardware acceleration. That power does, of course, come at a cost to the company and means that this is also the first time it has had to introduce liquid cooling solutions in its data centers.
As with the previous iterations of its TPUs, Google's primary goal with TPU 3.0 is to enable the build-out of new A.I. and machine learning models at an accelerated rate. Moreover, this takes away some of the burdens of building ML, such as the trial and error process for building custom hardware from the ground up. Google has said it will be utilizing some of its new TPUs for new A.I. features being brought to its own products such as Google Photos. However, TPU version 3 will also be open to developers. As of this writing, there hasn't been any information given as to the pricing Google will ultimately attach to the use of its new TPUs. The company also hasn't revealed the exact specifications for its new TPU Pods, such as how many TPUs will be included in an array or how many flops are generated by each individual TPU.
Accessing the TPUs via Google's Cloud Platform currently costs $6.50 per TPU per hour in the U.S. but there's no way to know if Google plans on changing pricing at all. With that said, Google I/O 2018 runs through May 10, so it wouldn't be surprising for the company to go more in-depth over the next few days and weeks. The TPUs were announced both via the search giant's official Twitter and during CEO Sundar Pichai's opening keynote speech.