In a bid to lure more companies to run their artificial intelligence projects on its cloud platform, Amazon Web Services Inc. today launched a family of AI-optimized instances described as the most powerful of their kind on the market.

The new P3 series is based on Nvidia Corp.’s latest-generation Volta graphics processing unit, called Tesla V100, beating Microsoft Corp.’s and Google LLC’s cloud units to the market. Unveiled in May, the chip packs 21.1 billion transistors on a die roughly the size of an Apple Watch face. These transistors form over 5,700 processing cores of which 640 are so-called Tensor Cores, specialized circuits fine-tuned for running AI models.

Amazon’s new P3 instance series comes in three sizes that provide one, four and eight V100s per virtual machine. The chips in the two largest configurations are connected using an Nvidia-developed technology called NVLink that allows them to exchange data for processing much faster than traditional alternatives allow. They’re supported by up to 64 vCPUs based on a modified version of Intel Corp.’s Xeon E5-2686v4 central processing unit.

All this silicon enables the largest P3 instance to provide up to a petaflop of performance under certain circumstances. That’s one followed by 15 zeros’ worth of calculations a second. In more practical terms, the series is about 14 times faster than AWS’ previous-generation P2 machines.

Matt Garman, an executive with the cloud giant, said in a statement that the latter instances already account for “most of the machine learning done in the cloud today.” That means the launch of the P3 series should only cement AWS’s lead. This is especially true since its two top rivals, Google Inc. and Microsoft Corp., don’t yet support Nvidia’s V100 chips in their respective cloud platforms.

“Once again AWS and Nvidia have taken the lead, set the price and the rest of the industry can follow,” Karl Freund, Moor Insights & Strategy’s consulting lead for high-performance computing and deep learning, told SiliconANGLE.

Companies can get started with the new P3 instances in one of two ways. They may set them up manually or use one of the two preconfigured Amazon Machine Images available for the series, which include a ready-to-use selection of popular AI frameworks. The first AMI was developed by AWS, while the other comes courtesy of Nvidia, which introduced the offering in a separate announcement.

Nvidia said AWS customers are the first to be able to use Nvidia’s AI Cloud Container Registry, a stack of software that includes the common open-source deep learning frameworks such as TensorFlow, Caffe, CNTK and Torch.

“This will unleash a lot of capabilities developers wanted,” Jim McHugh, vice president and general manager of enterprise systems at Nvidia. “The idea is to reach as many users as possible, democratizing deep learning.”

Although AWS is the first to make Nvidia’s latest technology available, McHugh said other cloud providers are expected to do so as well. He didn’t provide a time frame, however.

One of the first adopters of the P3 series is Schrödinger LLC, a New York-based chemical simulation provider. The company claims that the new instances allow it to perform four times as many simulations in a day as the previous-generation P2 machines.

Nvidia provided a video on how its GPU Cloud works on AWS:

With reporting from Robert Hof

Image: Robert Hof