MIT-Neural-Chip_0

MIT Researchers Unveil Computer Chip That Can Power Mobile AI

February 8, 2016 - Written By Muni Perez

Artificial Intelligence has been around for decades, but only in the past few years regular users are having direct contact with it and the technology is being used on simple daily things such as Google learning what we like to serve targeted ads, or Facebook algorithms so we see what matters on our timeline, and so on. You may wonder why it took so long for AI to come out of science fiction and land on Google Now, and why your phone can’t really tell the difference between a pear and an apple, and this is because machine learning needs a lot of energy and processing power. Last year Qualcomm showcased a machine learning platform embedded on a Snapdragon processor that could bring the power of AI to a smartphone, and MIT researchers have just unveiled a new special chip that is even more promising.

Called Eyeriss and unveiled during the International Solid-State Circuits Conference in San Francisco, CA, the chip is composed¬†of 168 cores that have their own memory, allowing it to work without the need to communicate with a central memory bank, saving time and energy. Additionally, the data to be processed is compressed before entering the cores, and a special circuit gives each core the maximum amount of work they can handle, and each core can directly communicate with their neighbor, meaning data don’t need to go through a central memory bank and can be shared locally, instead. The result is that Eyeriss consumes one-tenth of the energy needed by a regular mobile GPU, making the way for future installation on mobile devices.

Eyeriss is partially funded by DARPA, the DoD agency responsible for the development of emerging technologies for the military, and is a comeback from research on neural networks (which tries to imitate how the human brain works) conducted in the 70s and scrapped due to the large amount of power it needed with the technology available at that time. There’s a wide range of applications for machine learning capability on a smartphone and it would significantly reduce the need for an app to communicate with a server in the cloud, also reducing the need for internet connectivity for some apps to work. Looking away from consumer technology, Eyeriss could also be used in a wide range of autonomous cars, robots, drones and all sorts of applications that required computer learning and processing. There’s no¬†timeline of when the technology will be commercially available, thought it should take at least a few years before we see it, who knows on the Galaxy S12, or even Skynet.