In a new beta program, Google is making its custom machine-learning optimized processing units available to the public for use in their own applications through Google Cloud.
Google created its own processors, called the Tenor Processing Units (TPU), in order to accelerate machine learning by optimizing the processor for the type of processing that machine learning algorithms deploy. By creating these specialized units, Google can significantly increase the speed at which machine learning applications run.
In short, Google wants to facilitate smarter, faster artificial intelligence made possible by more powerful processing units in the cloud.
TPUs are unique as each unit contains four ASIC processing chips and boast a floating-point performance of up to 180 TFlops on 64GB of memory.
Application-specific Integrated Circuit (ASIC) processors are currently in use in data centers and homes around the world for Bitcoin mining as they handle the types of equations bitcoin miners are tasked with solving at a much faster and more efficient pace than traditional CPUs and GPUs.
The downside of using ASIC chips is that they’re only really good for a specific task, and not for general processing. You wouldn’t be able to run a server off an ASIC unit, but you could give your machine learning process a boost with one.
Turning TPUs Over to the Public
Now, some users can take advantage of Google Cloud’s limited supply for their own applications. Initial costs for running one of these second-generation TPUs is $6.50 per Cloud TPU per hour, billed by the second.
This cost doesn’t come without some merit. The first-generation TPU was clocked faster than CPUs and GPUs for calculating AI data by a factor of 30.
What TPUs aren’t good for is general processing. They’re powered by ASIC chips, which are very specifically aligned to calculate equations.
Currently, Google Cloud customers can rent individual TPU boards, but plans are in place to expand this offering to multiple connected boards in a product Google calls TPU Pods.
Competition on the Horizon
Google isn’t the only company creating its own AI-optimized chips. Amazon is in the process of developing one to improve its Alexa product and to offer to its AWS customers. The aim for this project is to speed up response time for Alexa-enabled devices by offering onboard processing that cuts down on the latency that comes with sending voice data to the cloud, having it processed remotely, then await a response at the device.
Machine learning and artificial intelligence are an emerging practice right now. Companies like Google, Microsoft, and Amazon are embracing it for their cloud customers as companies look for ways to establish the infrastructure they need to create their own AI applications.
Having processing units specifically designed to accelerate machine learning is another step in that evolution.