x
Cloud Native/NFV

Google's TPU Chips Beef Up Machine Learning

Google is offering more powerful hardware for its machine learning cloud customers as it launches its next generation of Tensor Processing Unit (TPU) chips through its Google Compute Engine.

The company announced these second-generation TPUs at Google I/O this week, and Jeff Dean, a Google Senior Fellow, and Urs Hölzle, senior vice president for technical infrastructure, wrote a detailed blog post on May 17 about these chips.

Officially called Cloud TPUs, these chips can work with Intel's Skylake processors, as well as Nvidia's GPUs. The new generation of TPUs offer 180 teraflops of floating-point performance, and Google has designed these chips so that they can be stacked in what the company calls a TPU pod, which houses 64 TPUs and can provide up to 11.5 petaflops of compute power.

Taken together, these TPUs are designed to accelerate machine learning, while giving customers access to the power of the Google's cloud, which can lower the barrier to entry for many businesses trying to build machine learning and artificial intelligence (AI) applications.

A TPU pod (Source: Google)
A TPU pod (Source: Google)

As Dean and Hölzle wrote in their blog:

"Our goal is to help you build the best possible machine learning systems from top to bottom. While Cloud TPUs will benefit many ML applications, we remain committed to offering a wide range of hardware on Google Cloud so you can choose the accelerators that best fit your particular use case at any given time."

Google has made machine learning and AI an essential building block of new products, including using the technology to improve search results, as well as in the development of its DeepMind Alpha Go program.


M&A activity is turning the cloud upside down. Find out what you need to know in our special report: Mergers, Acquisitions & IPOs Are Rocking the Cloud.


In addition, Google is making 1,000 Cloud TPUs available to researchers through its TensorFlow Research Cloud for free. The one catch is that anyone conducting machine learning or AI research would have to share their findings in a scientific journal or make them available through open source.

Google has been busy with cloud announcements this week. It pushed its Cloud Spanner database into general availability and released its Cloud IoT Core, a new service designed for businesses to help them manage all the data collected through Internet of Things devices. (See Google Cloud Spanner Hits General Availability.)

Related posts:

— Scott Ferguson, Editor, Enterprise Cloud News. Follow him on Twitter @sferguson_LR.

mhhfive 6/6/2017 | 4:20:48 PM
Re: TPUs vs CPUs/GPUs > "I've been reading up on Google's TPU research..."

I wouldn't bet against Google, either. Intel is in a tough spot, for sure. But it also owns a lot of IP related to making chips with nanometer-size features.... 
mhhfive 6/6/2017 | 4:19:12 PM
Re: Who has the best machine learning? Google's openness is a huge strength for attracting developers and gaining usage among academics. I wonder, though, how much Google will be hindered by not having a division that can make cutting edge chips? Intel and IBM have chip making capabilities with features as small as 5nm. Google is "playing" with quantum computing.. but is that really where machine learning is going? Is Google betting on being able to contract out sub-5nm chip designs to Intel and IBM? 
kq4ym 5/25/2017 | 3:55:36 PM
Re: Who has the best machine learning? With Google offering the 1,000 TPUs free with the mandate the user publish the results, it's probably a good bet that will work out well for all parties. Google will be working very hard with customers to enable them " to accelerate machine learning, while giving customers access to the power of the Google's cloud," and keeping a great competitive edge along the way.
Joe Stanganelli 5/23/2017 | 10:55:37 PM
Re: Who has the best machine learning? Google's pretty well diversified, but yes, the cloud is a big revenue stream -- especially in the life sciences...

...and, as it so happens, I'm at Bio-IT World Conference and Expo this week, where Google Cloud Platform is represented and exhibiting.  This stuff is huge in the life-sciences/healthcare/bioinformatics sector because of how big genomics and medical-imaging data are -- and how unwieldy.
Joe Stanganelli 5/22/2017 | 8:44:29 PM
TPUs vs CPUs/GPUs I've been reading up on Google's TPU research (including the original paper), and I'm so far convinced that the mega-giant is going to give Intel and those folks a run for their money with this new technology.
Michelle 5/20/2017 | 10:39:06 PM
Re: Who has the best machine learning? @mhh I think research in both fields will be used. It'll be scary and fascinating at the same time.
danielcawrey 5/20/2017 | 7:11:19 PM
Re: Who has the best machine learning? Google's focus these days is not just selling Gmail to enterprise, it's Google Cloud Platform. This is where the company is ultimately going to make its money. Apps and Gmail are just going to eventually be a small piece of the puzzle. 
mhhfive 5/19/2017 | 9:28:12 AM
Re: Who has the best machine learning? Does AI that beats humans at Jeopardy sound better than AI that can beat humans at Go? I suppose Jeopardy sounds more practical and possibly applicable to other more real world business tasks?
Scott_Ferguson 5/19/2017 | 9:12:47 AM
Re: Who has the best machine learning? @mhhfive: I'm not sure how you would compare what Google offers vs. what the other guys do. I would guess that a lot of that would come if/when people publish research based on what Google is offering and what kind of developments they are making. 
mhhfive 5/18/2017 | 6:28:11 PM
Who has the best machine learning? How will end-users be able to compare the machine learning capabilities of Google, IBM, MSFT, etc..? It'll be very difficult to make apple-to-apple comparisons (ahem, no pun intended, AAPL)....

Google seems to have staked out a space where it has a unique combination of language control *and* hardware integration. That sounds very Apple-esque. But Google sounds like a DIY model compared to IBM's consulting or partnership model. 
HOME
Sign In
SEARCH
CLOSE
MORE
CLOSE