MIT scientists have developed a new purpose-specific chip that will allow the processing of large datasets directly into the same electronic devices in the future, without the need to rely on neural networks available in the cloud. This design will help the neural networks can be implanted in mobile phones and other electronic devices in the future.
A chip that will facilitate the integration of local neural networks in electronic devices
From the MIT point out that this new chip will save your energy consumption by up to 95%. And even offer a speed of calculation of neural network three to seven times compared to the predecessors. Simplified chip design has been sought, in which the transfer of data, back and forth, from one part of the chip to another and vice versa, is eliminated, which is where the highest energy consumption is generated.
And to do this, the calculation function has been integrated as a specific operation within the memory, to avoid continuous data transfers within the same chip. In this way, future devices could perform functions of voice recognition, face recognition, among other operations, locally.
The design project for the new chip has been led by Avishek Biswas, an undergraduate student at MIT in electrical and computer engineering, and is described in a document to be presented this week at the International Conference on Solid-State Circuits.