Beijing time on the afternoon of February 28, according to MIT News reported that MIT researchers have developed a high-performance chip can be used for neural network computing, The chip can process up to seven times the speed of other processors while requiring 94-95% less power than other chips, and in the future such chips will likely be used on mobile devices running neural networks or Internet of things equipment.
Avishek Biswas, a graduate student in electrical engineering and computing science at MIT, a leader in this project development, said: 'The general processor operating mode is such that some of the chips Part of the memory is placed, it will move data back and forth in these memories as the calculations take place. Since machine learning algorithms require a lot of computational effort, they consume a lot of energy when moving data back and forth, but in fact these algorithms The calculations made can be reduced to a specific operation called a dot product. Our idea is whether we can deploy this dot product function in memory, eliminating the need for constant movement These data? '
The chip converts the input of the node into a voltage, which is then converted to digital form for storage and further processing. This approach allows the chip to calculate the dot product of 16 nodes in a single step without the need to move data between the memory and the processor. MIT News believes this approach is closer to working in the human brain the way.
Biswos will be elaborating on how the chip works in a paper that will be presented during the International Solid State Circuits Conference, along with his dissertation co-authors, MIT Engineering College Dean Anantha Chandrakasan, and Vannevar Bush, professor of electrical engineering and computer science at MIT.
Last December, SensibleVision CEO George Brostoff published a guest post at the Biometric Update that demonstrated the potential of custom processors for mobile devices Since then, FWDNXT has also announced that they will develop low-power processors that use deep neural networks for image recognition and classification, and in addition ARM also announced that it will develop tools for machine learning and object recognition chip.
