IBM S822LC server Powers Tesla P100 deep learning
Nvidia CEO Jen-Hsun Huang has been growing his company into something greater than a graphics chip vendor for some time now. A new partnership between IBM and Nvidia to make computer systems for deep-learning applications reinforces the idea that the graphics giant's technology can be put to more interesting uses than rendering video games at blistering frame rates. The fruit of IBM and Nvidia's collaboration is the Power System S822LC, a server that joins two eight or ten-core IBM Power8 CPUs with up to four Nvidia Tesla P100 GPUs over Nvidia's NVLink architecture.
The companies claim the new system delivers "2X performance over similar servers" when performing certain deep-learning tasks. According to Nvidia and IBM, the 822LC's performance is similar to that of x86 servers with eight of Nvidia's previous-generation Tesla M40 GPUs. The servers' Nvidia Tesla P100 chips are based on Pascal GP100 silicon with 16GB of on-package HBM2 memory. Each GPU is capable of 21.2 TFLOPs of the half-precision arithmetic often used in deep learning research. Nvidia's NVLink interconnect technology provides up to 160 GB/s bandwidth to each GPU. The IBM Power8 CPUs are available in two configurations: pairs of eight-core CPUs at 3.25GHz or doublets of 2.86GHz ten-core processors.
IBM will offer a software toolkit called PowerAI that includes binary distributions of several deep learning tools optimized for the new system, including Berkeley Vision's and Learning Center's Caffe frameworks, the scientific deep-learning platform Torch, and the Python deep-learning library Theano. PowerAI is designed to run on a single server or scale to deployments of thousands of machines.
Some of the partnerships' initial customers include the Human Brain Project, HPC cloud hosting provider Nimbix, and Turkish cloud supercomputing center SC3 Electronics. IBM and Nvidia did not offer availability or pricing information, but rest assured, the prices are high.
In related news, the company recently highlighted BabbyCam in its blog, the brain-child of one of its deep-learning engineers. BabbyCam is a baby monitor that can observe an infant and use image processing to determine if it's sleeping, awake, crying, or has its face covered. The system was developed using Nvidia Tesla hardware and the Caffe framework, and can automatically send email or text messages to parents when any of these conditions is detected.