AWS AI Blog

Announcing New AWS Deep Learning AMI for Amazon EC2 P3 Instances

We’re pleased to announce a new set of AWS Deep Learning AMIs, which come pre-installed with deep learning frameworks optimized for the NVIDIA Volta V100 GPUs in the new Amazon EC2 P3 instance family.

The new P3 instances are perfect for deep learning, with eight NVIDIA Volta GPUs available in a single p3.16xlarge instance, which is capable of 125 trillion single-precision floating-point operations per second. This dramatically reduces the training time for sophisticated deep learning models, from days to just hours.

Deep Learning AMIs: Now with CUDA 9 for Volta

The new AMIs are pre-installed and configured with CUDA 9 for both Ubuntu and Amazon Linux, along with the other GPU drivers to take advantage of the speed of Volta on P3, including cuDNN 7.0, NCCL 2.0.5 and the NVIDIA Driver 384.81. 

Deep learning frameworks optimized for Volta

In addition to the latest drivers, the AMI also includes popular frameworks which have been optimized for P3 and Volta. Volta is brand new, and so this first release of the new CUDA 9 AMIs includes frameworks which were built from source, using the latest available versions. It’s still super-early for Volta, and so while we expect the frameworks to continue to evolve in terms of stability and performance over time, these latest versions will provide significant performance improvements in training time. That said, some of these builds are from the bleeding edge, so test before you put them into production. You can help them all get better by sending feedback to the open source projects as you run your experiments on P3.

Apache MXNet v0.12 RC1: Training convolutional neural networks is up to 3.5 times faster than the Pascal GPUs when using float16. Full release notes for MXNet version 0.12 are available here.

Caffe2 v0.8.1: With FP16 support, Caffe2 allows developers using NVIDIA Tesla V100 GPUs to maximize the performance of their deep learning workloads.

TensorFlow (master): We include a version of TensorFlow built from the master as of 11am on 24th October, 2017 (commit: 5bef42), and included 8 patches from NVIDIA for Volta support. In our initial testing, this build shows 3 times the training performance on a p3.16xlarge compared to the latest official release (TensorFlow 1.3), on p2.16xlarge.

Gluon: As with previous versions, these AMIs includes Gluon, a new open source deep learning interface which allows developers to easily and quickly build machine learning models, without compromising training speed. You can read more about Gluon in our launch announcement, and get started with over 50 notebooks with sample code. 

Selecting the right AMI for the job

Getting started with the AWS Deep Learning AMI is simple. You can start with just one click from the AWS Marketplace or follow this step-by-step guide to get started with your first notebook.

The CUDA 9 AMI is available in the following versions:

We’ll add additional frameworks as they become available for Volta. The CUDA 8 AMIs are still maintained and available, and include the latest stable, official “point” releases for TensorFlow, MXNet, Caffe, Caffe 2, Microsoft Cognitive Toolkit (CNTK) and Keras. The CUDA 8 AMI is available in the following versions:

Join us at AWS Re:Invent 2017

If you’re interested in learning more about deep learning, don’t forget to also check out our Guide to Machine Learning at re:Invent, November 27 to December 1, 2017. At the re:Invent conference you can explore over 50 machine learning sessions, workshops, and labs as well as attend our inaugural Deep Learning Summit for perspectives on the future of deep learning.