Fujitsu Graphics Processing Unit (GPU)

Fujitsu Graphics Processing Unit (GPU) is the software technology integrating several graphic processors for obtaining advantages of high-speed deep training.

On August 18, 2016 Fujitsu announced development of a software technology which integrates several graphic processors (GPU, Graphics Processing Unit) for obtaining advantages of high-speed deep training.

Deep training – the technology showing higher level of accuracy of recognition in comparison with the previous developments. Ensuring high accuracy of recognition requires the permanent analysis of data bulks. Graphic processors which are suitable for high-speed transactions in comparison with normal central processors (CPU) better were for this purpose used.

Differences between technologies, (2016)

As a rule, acceleration of deep training requires several computers with powerful GPU, united in network and working in parallel. The lack of this technique - at consolidation more than 10 computers increases data exchange time between computers and it is more difficult for specialists to achieve effect of parallelism.

Fujitsu processed technology of a parallelization which provides effective and fast data exchange between computers. The technology was tested within the platform of deep training Caffe. Within the test measuring training time using AlexNet 1 and network of computers with 64 GPU development showed the training speed which is 27 times higher, than during the operation of one computer with one GPU. In comparison with results without use of technology, the speed of training grew by 46% for network with 16 GPU and for 71% for 64 GPU.

Action of a paralellizm, (2016)

Using technology, problems of machine learning for which accomplishment one month was required earlier can be executed approximately in a day in a parallel operation mode 64 GPU.

Within development and deployment of technology, Fujitsu it was succeeded to increase data processing rate at deep training:

Technology of installation of a priority of data exchange

This technology automatically controls a data exchange priority in order that this, necessary to start the next session of training, exchanged between several computers beforehand. In the existing technologies processing of tasks of data exchange of the first level which are necessary to start the following learning process is executed by the last that increases a delay. In the provided technology, due to accomplishment of processing of tasks of data exchange of the first level during processing of tasks of data exchange of the second level, waiting time is reduced.

The technology of processing optimizing working activities according to data size

For data processing at which results of transactions are transferred to all computers and initial amount of data small, each computer exchanges data and then executes the same transaction that allows to eliminate time necessary for transfer and obtaining results. When amount of data big, processing is distributed, and results of processing are transferred to other computers for the subsequent transactions. Due to the automatic choice of an optimal method on the basis of data size this technology increases the general speed of work.

By means of the created technology it is possible to reduce time of research and development in the field of deep training, including development of unique models of neural network for automated management of robots, cars and so forth Besides, advantages of development can be used in the field of finance and health care, including tasks of classification of pathologies and forecasting of stock exchange rates. Fujitsu is going to begin to take in 2016 commercial benefit from development within the project of artificial intelligence Human Centric AI Zinrai.