Utilized with Nvidia NVLink™ and GPUDL Library Optimized for IBM Power Architecture
IBM PowerAI Software Toolkit Achieves 2x Performance with Caffe-based AlexNet

Nvidia (CEO Jensen Huang) announced that it is collaborating with IBM to develop new deep learning tools optimized for the latest technologies from both IBM and Nvidia. This collaboration aims to rapidly train computers to think and learn in a manner more similar to humans.

Deep learning is a rapidly growing machine learning approach that extracts information by segmenting data into millions of pieces and detecting the most critical aspects of the data to rank them according to importance. Supported by public endorsements from leading consumer web and mobile application companies, deep learning is rapidly being adopted in traditional enterprises as well.

Deep learning and AI are widely used across various industries. In banking, it is being used to detect financial fraud through facial recognition capabilities. In the automotive industry, it is being applied to the development of autonomous vehicles. In the retail industry, companies are introducing computers that can better understand conversations and answer customer questions, with the goal of achieving complete automation of call centers.

<Nvidia Artificial Intelligence Homepage>

The newly announced deep learning software toolkit IBM PowerAI (Power AI) runs on recently released IBM servers for AI. These servers include Nvidia® NVLink™ (NVLink™) interconnect technology optimized for IBM Power architecture. This hardware and software solution delivers more than 2x the performance compared to similar servers through four GPUs running AlexNet with Caffe. An identical four GPU Power-based configuration running BVLC Caffe-based AlexNet surpasses the performance of an eight M40 GPU-based x86 configuration, making it the fastest commercial enterprise system platform in the world for two versions of major deep learning frameworks.

Caffe is a deep learning framework developed by Berkeley Vision and Learning Center (BVLC) and is recognized as the most popular deep learning community application in the technology industry. Caffe is one of five deep learning software frameworks available in the IBM PowerAI toolkit. This toolkit is designed to maximize the utilization of the Nvidia GPUDL library, which includes cuDNN, cuBLAS, and NCCL, as part of the Nvidia SDK that enables multi-GPU acceleration on IBM servers.

IBM PowerAI is designed to run on the highest performing servers in IBM's OpenPower LC (OpenPOWER LC) lineup. These servers include Power architecture and Nvidia NVLink™ (NVLink™) technology optimized for Nvidia's latest GPU technology. This new solution supports emerging AI approaches, particularly deep learning. IBM PowerAI continues to expand AI expertise in the enterprise sector by adopting multiple deep learning methods used to train IBM's cognitive solutions platform, Watson.

IBM PowerAI is provided at no additional cost to customers who purchase IBM Power S822LC for HPC servers. PowerAI is designed to run on a single S822LC server and can be scaled to large supercomputing clusters configured with tens, hundreds, or thousands of systems.

PowerAI is a binary distribution set of popular deep learning frameworks such as Caffe, Torch, and Theano. Additional distributions include IBM-Caffe and NV-Caffe, which are IBM and Nvidia versions of the Caffe deep learning framework respectively. IBM has optimized each distribution to take advantage of the Nvidia NVLink™ (NVLink™) interface included in the IBM Power S822LC for HPC servers and the benefits of the recently announced IBM POWER8 (POWER8) chip.