Nvidia's top-shelf GPUs have arrived on Amazon Web Services (AWS). The cloud provider announced this week the general availability of its P4d instances, powered by Nvidia's Ampere-based A100 GPUs.

The announcement makes AWS the last of the public cloud providers to adopt Nvidia's A100 GPUs with Google, Microsoft, and even Oracle announcing support for the GPUs earlier this summer.

"The P4d instance delivers AWS's highest performance, most cost-effective GPU-based performance for machine learning and high-performance computing applications," wrote Ian Buck, GM and VP of accelerated computing at Nvidia, in a blog post Monday.

AWS' new instances are tailored to the needs of data scientists training machine learning algorithms on large data sets. Each P4d instance features eight A100 GPUs, announced by Nvidia in April. However, customers using AWS UltraClusters can scale this up to more than 4,000 GPUs at a time using AWS's Elastic Fabric Adapter (EFA).

To feed each P4d instance with data, AWS is taking advantage of Nvidia's high-speed interconnect technologies including NVLink, NVSwitch, Nvidia Collective Communications Library, and GPU Direct remote direct memory access, which enables up to 400 Gb/s of networking throughput per instance.

According to Buck, the higher performance offered by the P4d — up to 2.5 petaFLOPs for a single instance — enables customers to train machine learning models up to 300% faster using 16-bit floating-point calculations and 600% faster when using Nvidia's 32-bit Tensor Float calculations.

While Amazon's P4d instances feature eight A100 GPUs and more than a terabyte of system memory, the cloud provider isn't using Nvidia's DGX platform announced back in April. Instead of the 128 thread AMD EPYC 2 CPUs used in the DGX A100, Amazon opted for Intel's Cascade Lake Xeon Scalable CPUs, with each instance featuring 96 virtual CPUs.

P4d instances can be spun up in Amazon's Elastic Container and Kubernetes services, AWS ParallelCluster, and SageMaker. They are now available in the U.S. east and west districts and are coming to additional regions in the near future.

GE Healthcare, Toyota Research Institute, Aon, Rad AI, OmniSci, and Zenotech are already leveraging the new instances or have announced plans to deploy their data science workloads on AWS P4d instances.

"Using the new P4d instances reduced processing time from days to hours," said Karley Yoder, VP and GM of artificial intelligence at GE Healthcare, in a statement.

Aon saw similar improvements in data processing. According to Peter Philips, CEO of Aon's PathWise division, AWS's P4d instances offered a twofold reduction in cost and 250% increase in speed. PathWise is a GPU-based risk management platform used by insurers, banks, and pension funds to hedge economic risks.