Chinese engineers have developed a hybrid integrated circuit that is capable of natively processing several different types of machine learning algorithms. They can do this at the same time.
Deep convolutional neural networks have occupied the concept of machine learning over the course of a decade, but it is only one of many varieties of it. For example, biological organisms are better imitated by a spiking neural networks. The two types operate on quite different principles.
Therefore, the assembly of a neuromorphic circuit capable of both types of operation is a major challenge. And this is precisely what has now succeeded the Chinese from several Beijing universities and institutes.
The Tianjic chip consists of 156 FCore cores and runs at 300 megahertz, with 40,000 nodes and ten million synapses. There are 256 computational units in one FCore core, capable of machine-level representation of both nodes of convolution and shock neural networks. Compared to the emulation on the Titan XP graphics card, it depends on the algorithm, from 1.6-to a hundred times faster, and the power consumption from 12 to 10,000 more efficient in electrical consumption.
Source of shared Link