Cloud

Your AI workloads in the cloud: Google finally releases TPU beta

The Tensor Processing Units could help artificial intelligence experts train and deploy their machine learning models more quickly.

Building a slide deck, pitch, or presentation? Here are the big takeaways:
  • Google has released its Cloud TPU machine learning accelerators in beta, which could help speed the training of machine learning models.
  • Enterprises looking to increase their focus on machine learning and AI should consider the Google Cloud Platform as a viable option for their workloads.

Google released its Tensor Processing Units (TPUs) in beta on the Google Cloud Platform on Monday, providing a strong option for enterprises to boost their efforts in artificial intelligence (AI) and machine learning.

Announced via a blog post, the TPUs will go a long way toward speeding the training of machine learning models—creating models that can be trained overnight instead of over days or weeks. Google first announced its work with TPUs some years ago, but is just now releasing them to be used by its cloud customers.

The release of the TPUs adds to the growing number of compute options available to companies that want to do serious machine learning and AI work in the cloud. In October 2017, Amazon Web Services (AWS) made the NVIDIA Tesla V100 GPUs available in EC2 among a host of other machine learning tools, and Microsoft also made the same GPUs available in late 2017 for HPC and AI workloads.

SEE: IT leader's guide to deep learning (Tech Pro Research)

Given the growing importance of AI and machine learning, the ability to process the workloads and train these models will become table stakes among the top cloud providers. Google has made a name for itself as one of the premier names in machine learning and deep learning, and the availability of TPUs should be a serious consideration for cloud customers looking for a place to run their AI and machine learning workloads.

In its blog post, Google describes Cloud TPUs as a "family of Google-designed hardware accelerators that are optimized to speed up and scale up specific ML workloads programmed with TensorFlow." These accelerators have been the force behind parts of Google data centers since 2015, as noted by our sister site ZDNet, and offer 180 teraflops of floating-point performance on a single board, the post said. At a recent Google I/O event, CEO Sundar Pichai said that Google was rethinking its computational architecture to build "AI-first data centers."

Instead of having to share a cluster, data scientists get access to a network-attached Cloud TPU through a Google Compute Engine VM, the post said. They can control and customize those to meet the needs of their workload.

In addition to offering TPUs, Google Cloud Platform also offers access to CPUs like the Intel Skylake series, and GPUs such as the aforementioned NVIDIA Tesla V100. Cloud TPU billing is calculated by the second at a rate of $6.50 per Cloud TPU, per hour, the post said.

In a separate post, Google also announced that GPUs in Kubernetes Engine are in beta. These could also provide a boost to similar efforts in machine learning and image processing.

Also see

tpu.jpg
Image: Google

About Conner Forrest

Conner Forrest is a Senior Editor for TechRepublic. He covers enterprise technology and is interested in the convergence of tech and culture.

Editor's Picks

Free Newsletters, In your Inbox