Not far off is the day when mankind rebelled against the machine go back in time a cyborg assassin, because the team of experts from the Massachusetts Institute of Technology has developed a processor that greatly accelerates data processing and pattern recognition neural networks. Their development in the relatively near future may give wearable electronics, the ability to think and make decisions.
The processor codenamed Eyeriss was presented at the international conference of microelectronics manufacturers in San Francisco. Eyeriss was originally developed to facilitate deep learning. processor creators claim that their chip is able to exceed ten in a data rate of some modern GPU. In other words, this technology very soon it will be possible to adapt and for portable devices such as tablets and smartphones.
"The use of deep learning process is very important in many different areas. This set of algorithms allows for object recognition, speech, human faces. Today neural networks are very complex and are able to work only on powerful GPU. Just imagine that you can get all this functionality to your smartphone, "- says associate professor of electrical engineering at MIT Vivienne See.
The impressive increase in the processing speed has been achieved at the expense of a very cunning processor design. First, instead of using a common pool of memory, each of the 168 nuclear Eyeriss has its own discrete cache. This will get rid of unnecessary data transfer within the system. Second, to reduce the data rate, the processor first compresses information using a specific circuit, and then decompresses it to the destination.
But most importantly, Eyeriss processor uses a special cunning scheme by which the data processing can be configured so that the data is distributed between the cores more efficient way. Evenly distribute the load, CPU time per unit which handles more data than some modern GPU. However, there is nothing about the specific application of the new chip at the conference has not been told. We can only patiently wait for new information.