Development of a very high-speed, low power computer system for Deep Learning at Perferred Networks

Deep learning is a powerful and flexible solution for dealing with complicated phenomena or dynamic environments that are difficult to handle with conventional rule-based programming. PFN’s strength is to combine a profound knowledge of deep learning together with expertise in various other fields to develop state-of-the-art technologies. We have been engaged in challenges beyond boundaries ― from developing the core deep learning technology Chainer, to building large-scale compute clusters, to exploring a wide variety of domains (including robotics, life science, and others) ― with the aim of contributing to the world with technologies that only PFN can realize. To speed up the training phase in deep learning, PFN is currently developing the MN-Core chip, which is dedicated and optimized for performing matrix operations, a process characteristic of deep learning. MN-Core is expected to achieve a world top-class performance per watt of 1 TFLOPS/W (half precision). Today, floating-point operations per second per watt is one of the most important benchmarks to consider when developing a chip. By focusing on minimal functionalities, the dedicated chip can boost effective performance in deep learning as well as bringing down costs