Under the leadership of CEO Jen-Hsun Huang, Nvidia has grown into a detached graphics chip manufacturing companies. It reached a new IBM "deep learning to create computing system" new cooperation, and introduced the Power System S822LC. With Nvidia's NVLink architecture, the server is connected to a dual eight-core (or ten-core) IBM Power8 CPU and up to four Nvidia Tesla P100 GPUs.
The company said that in the implementation of some of the depth of learning tasks, the new system can bring more similar 2x performance.
According to Nvidia and IBM, 822LC performance, and Nvidia previous generation 8-way Tesla M40 GPU (x86 server) is similar. 822LC equipped with Nvidia Tesla P100 chip, based on Pascal GP100 architecture to build, supplemented by 16GB HBM2 memory.
In-depth study of the commonly used semi-precision calculation, the computing power of each GPU reached 21.2 T-Flops. In addition, Nvidia's NVLink interconnect technology provides up to 160GB / s per GPU.
IBMPower8 CPUThere are two configurations, eight cores @ 3.25GHz, or ten cores @ 2.86GHz.
IBM will provide a software toolkit called PowerAI (for a variety of application-depth learning optimization), including Berkeley Vision and Learning CenterCaffeFramework, scientific depth learning platformTorch, And the Python Deep Learning LibraryTheano.
Power AI can run on a single server, or on thousands of machines.
IBM's initial partnership customers include the Human Brain Project Team (Human Brain Project), A provider of high performance computing cloud hostingNimbix, As well as the Turkish cloud supercomputing centerSC3 Electronics.
IBM and Nvidia did not disclose the exact pricing and launch time, but we guess will not be too cheap.