Six years ago IBM as part of the project for DARPA SyNAPSE agency task began to create a semiconductor analog of the human brain. In 2014 it was released the second version of the development, which was called TrueNorth. This 28-nanometer chip production of Samsung one million digital "neurons", each of which is connected to 256 digital "synapses" (total - 256 million synapses). Generally TrueNorth processor contains 5.4 billion transistors. This spring, the first computer-based TrueNorth was sold - NS16e system consisting of 16 processors TrueNorth related likeness of neural networks. Computer for research in the field of AI has acquired Livermore National Laboratory. Lawrence.
As reported recently IBM company , the researchers were able to create a unique algorithm to run on TrueNorth platform, which will allow to organize a high-performance and compact self-learning system. For example, if a typical problem of pattern recognition on the basis of deep machine learning is performed on the basis of calculations of the accelerator with the consumption of about 150 watts, the TrueNorth processor using the new algorithm is able to accomplish this with the meager consumption, which is comparable to the smartphone battery consumption within a few days.
New algorithm using TrueNorth processor was able to classify the visual images at a rate of 1200 to 2600 frames per second with the consumption of 25 to 275 mW. Thus, work efficiency was 6000 FPS / Watt. In fact, IBM TrueNorth processor is capable of real-time process frames concurrently with 50-100 cameras at 24 frames per second in a matrix of 32 x 32 pixels. And all this happens again, against the background of a very small consumption. That is the whole meaning of the simulation processor of the brain where the data is processed locally.
Now a few words about the new algorithm. Details of the development described in the scientific journal Proceedings of the National Academy of Sciences ( PNAS). The essence of the new method is that to run on the processor TrueNorth adapted the so-called convolutional neural network. Prior to this, the neural network (in general, not only convolutional) created on the basis of processors with the classical von Neumann architecture.. Chip TrueNorth demonstrated the ability to be an effective platform for systems associated with deep machine learning. Related Products :
|