At Google I/O last month, Google announced a new program called the TensorFlow Research Cloud, giving artificial intelligence researchers and data scientists access to a cluster of 1,000 Cloud TPUs to refine their machine-learning models with Google's purpose-built machine-learning chip.
Continue Reading Below
Now, Google is gearing up to make Cloud TPUs available for businesses through Google Cloud Platform.
Cloud Tensor Processing Units (TPUs) are part of a trend toward AI-specific processors, and for Google in particular these cloud-based TPUs are the underlying compute element driving a top-to-bottom AI rewrite fundamentally redefining how Google's apps, infrastructure, software, and services function by building intelligence in from the ground up.
At the Wired Business Conference in New York City today, Google SVP of Technical Infrastructure Urs Hölzle broke down the machine-learning power Google is democratizing with Cloud TPUs, and what businesses will soon be able to do with it.
"When you create these [neural networks] you end up doing a lot of math, but it's a specialized kind of math. So if you build a special-purpose chip, you can do it much more efficiently," said Hölzle. "If you try to drive a regular car there's this blinding amount of attributes. If you look at a race car, it's made for a specific situation. What we built is kind of a drag race car. Go straight and go as fast as you can. The TPU we built does machine-learning computation and nothing else."
Hölzle said there's no specific date on making Cloud TPUs available to businesses via Google Cloud, but that it's coming "soon." Google is using TPUs in everything from optimizing usage in its data centers to suggesting auto replies in Gmail. According to Hölzle, the AI-specific processing unit has 11 petaflops of capacity and can handle 180 trillion transactions per second related specifically to creating and training machine-learning models. For businesses, he explained how Cloud TPUs combined with the TensorFlow developer toolkit will allow businesses to develop machine-learning algorithms and applications for a wide variety of devices and use cases.
"We're making this available soon on Google Cloud Platform. TPU is the solution for compute power and our toolchains are getting better with Cloud ML and TensorFlow to help you build these neural network models," said Hölzle. "It's also portable across different platforms. Run it on a Mac, run it in a data center, on an Android phone, and all kinds of hardware."
What Businesses Can Do With Cloud TPUs
PCMag caught up with Hölzle after his session to talk about what businesses will actually be able to do with this technology that they couldn't do before. He gave several examples of what businesses are already doing with Cloud TPUs and TensorFlow, one of which was a Japanese e-commerce site.
"There are many customers successfully using it to do all kinds of things. We had a Japanese used car site where they used TensorFlow and Cloud ML to recognize the pictures their agents took of used cars—the model, make, year, condition and photos of the front, back, and interior—all sorts of pictures for a nice presentation on their website," said Hölzle. "The model fills out all the default info on the listing and can suggest a price range based on the odometer and any scratches or damage. That's an application they created in a matter of months."
Another customer, which offered satellite topography imagery to customers, used machine learning to solve a problem they'd been working on for 20 years. Customers don't want clouds in their mapping data, according to Hölzle. TensorFlow improved the accuracy of the company's cloud removal by a factor of four over a period of six months. Hölzle said there are applications across finance, commerce, medicine (like what Google Brain is doing), consumer apps, and beyond.
"There are limits, it's not magic, but it's really exciting how many places it's applicable and in how many businesses it makes sense," said Hölzle. "We're aiming to be the cloud platform for machine learning and analytics. We're making it much more accessible to average companies because it works across so many circumstances, from AlphaGo and data center cooling optimization to image and speech recognition trained on the same neural network."
Cloud TPUs are just one part of Google's larger pipeline for helping businesses do this. Hölzle walked through what this machine-learning offering will look like, going step by step through the process businesses will experience in Google Cloud Platform when creating, training, and deploying ML models and applications.
"The pipeline is about going from idea to solution. One of the first things you need to think about is what data set you can use to learn from. If it's something with an image, you need images that tell you what you want the model to learn. We have some fully finished ML to use in your application with the APIs we provide," said Hölzle.
"Let's say you want to do fraud detection from credit card transactions, but you also have noisy data," Hölzle continued. "Maybe the labels aren't correct. Your existing systems say a charge is fraudulent and then a custome calls to complain saying they wanted to make that purchase. Part of that is joining the data sets and cleaning that data if needed."
So you choose a neural network, Google helps you do the training, and then validate the results. Hölzle explained that in Google's tool set, once you have a trained model you simply tell Google what you want to do with it, and they run the service out of the box. Hölzle said Google will also help you iterate on that model, calibrating it through Cloud TPUs to optimize your algorithm for higher accuracy. One caveat: all that iteration might cost you.
"It's not just the core ML step. Once you have a machine learning project, 10 percent of time is spent on ML and 90 percent is on data preparation, cleaning, interpreting results, and iterating to find better models," said Hölzle. "We have something in Cloud ML to automatically try out multiple models to find what works best. You may have to pay more because the training step is hundreds of thousands of times more compute power, but you get the optimal model and higher accuracy just by pushing a button and waiting four hours."