NVIDIA Corporation has long been promoting the idea of creating a trained cloud systems that allow end devices use a collective experience to perform certain operations, as well as decision-making. Similarly we have continuously improved the system autopilot electric Tesla: everything that happens with private vehicles accounted for the rest, after the exchange of information.
NVIDIA specialists, according to a press release on the website of MIT , have been implicated in the development of a 168-core processor, using the principle of autonomous neural network to work with artificial intelligence. Many cores are equipped with own memory, and they can communicate with each other directly, bypassing the system memory. All computational work required for pattern recognition and speech is performed locally without resorting to network resources. This allows both to ensure the functioning of the device in the absence of communication, and improve information security - the data will not be sent.
Experimental Eyeriss processor provides energy efficiency is ten times superior performance of advanced mobile graphics solutions. At the same time it provides a work of artificial intelligence algorithms on a device having a compact size - from smart phones to wearable electronics. On the base of Eyeriss it will be possible to create not only "smart devices", but also robots that have some degree of autonomy in decision-making. Financing processor development was carried out in part at the expense of the US defense department DARPA.