IBM and Nvidia increased the speed of deep learning neural networks

For neural networks - the future of mankind, some experts believe. Or, if not all of humanity, then a number of areas related to information technology. IBM has been working for a long time to create an effective cognitive system that can solve complex problems and help people find solutions to a number of problems. One of the important issues is the deep learning of neural networks.

This process should be fast enough, but at the same time effective. IBM has entered into a collaboration agreement with Nvidia, which is to develop IBM Power AI software. This is a platform running on a specialized IBM Power System S822LC server on IBM Power 8 processors. Such a hardware and software platform, according to experts, should provide fast, in-depth training for modern networks.

A feature of the Power System S822LC server is its Nvidia NV Link interface. This interface is a means of direct interaction between the CPU and the GPU. From such servers, the project participants made up a test neural network, Alex Net, which uses the Caffe framework from Berkeley Vision and Learning Center . In addition to Caffe, the IBM Power AI platform includes libraries such as cuDNN, cuBLAS and NCCL. Caffe is a widely used framework that is considered one of the most popular in the field of in-depth training of neural systems.

The developers claim that the performance of the new system is several times higher than the performance of other systems built on ordinary servers. Partners say that the performance of the 822LC server is equal to the performance of the X86 system architecture with eight previous-generation Nvidia graphics cards - the Tesla M40. In general, the S822LC server with NVLink can transfer data about five times faster than similar x86 systems.
')
Actually, this server is the first Power8 system that comes with NVLink. Prior to the release of the new data exchange between the central processor and the GPU was a bottleneck in the PCIe x16 Gen3 interface. Now all the data is in the common space of the system memory, which can be said to facilitate programming.

Power System S822LC Configuration

- IBM POWER8 processors, standard operating at 2.92 or 3.32 GHz, with 8 or 10 fully activated cores;
- Up to 1024 GB of RAM;
- Five PCIe Gen3 slots, two of which only support GPUs;
- Two hard drives with SATA interface for 1 TB, 7200 rpm, or two SSD SATA drives for 480 (or 960) GB;
- Two Compute Intensive Accelerator GPU K80;
- One front USB 2.0 port and one rear USB 3.0 port;
- Two power sources;
- 19-inch rack-mount execution 2U high;
- Operating Systems:
Ubuntu Server 14.04.5 LTS
Ubuntu Server 16.04.1 LTS
Red Hat Enterprise Linux (RHEL) 7.2



The Power AI, among other things, includes several depth learning tools optimized for the new system. For example, this is the Torch science platform and the Python library. In general, the Power AI platform can work both on one server and on thousands of machines, it can be easily scaled.

“Power AI will simplify in-depth training and facilitate work with other analytical technologies. Scientists and researchers will get an easily scalable platform that can be deployed fairly quickly, ”the developers said. They also said they consider the new platform the best in terms of creating AI applications. It easily allows you to create "intelligent" chat bots, or work with statistics to get processed analysis results in a short time.

Organizations and projects such as Human BrainProject , HPC cloud provider Nimbix , and the Turkish computing center of supercomputers SC3 Electronics expressed their interest in trying out a new platform.