MIT Researchers Develop Efficient Neural Chip To Perform Complex AI Tasks With Ease

Researchers from MIT have developed a new energy efficient neural chip that can run powerful AI algorithms locally, without depending on the internet for heavy data processing. The chip, called “Eyeriss”, is ten times as efficient as a mobile GPU, and could greatly increase the task-handling capability of mobile devices.

MIT-Neural-Chip

A neural network is a vast virtual network of information processors, modeled to emulate the processing technique of the human brain. They are used to execute powerful computational processes, and are used to run algorithms that are too heavy for a single low-power device to handle.

Neural networks are typically implemented on GPUs (Graphics Processing Units) due to the higher number of cores a GPU possesses. Today’s neural networks are highly advanced processing units, but due to their complexity, need to be implemented on high-power GPUs. This means that these networks need a vast array of supporting systems, and are hardly portable. Any device that needed to make use of the network had to send data over, and the processed data would be sent back to the device; all through the internet. The whole process has obvious drawbacks like wastage of time and unreliability of a wireless connection.

The complexity of a neural network comes from the fact that it is organized into layers, with each layer containing a number of nodes that process information. These nodes process data like they have been trained to, finding connections between and applying labels to the information provided. The information is processed layer by layer, and many of the nodes in a layer process the same information in multiple ways. This increases computing power, but also power consumption.

The challenge faced by the researchers at MIT was reducing the power consumption of their chip, without sacrificing the network’s computing power.

The Eyeriss has 168 cores, and each of these cores has its own memory unlike a single, shared memory bank found in other GPUs. Hence, the cores would not have to transmit data back and forth the memory bank, thus saving time and energy. This also facilitates communication between individual cores without the need for it to be routed through the main memory.

The chip also has a special purpose allocation circuit that assigns tasks to the cores. The allocation unit also transmits information as to which core is processing what data across the chip, maximizing the efficiency at which the chip works.

The coming of the Eyeriss draws in a new age in the development of localized AI capable of independent processing of complex data, minimizing its dependence on the internet. Localized processing also means that mobile platforms will be able to execute algorithms required for visual and speech recognition, and could be a huge step towards the development of increasingly autonomous robots.

The ability of the Eyeriss to be implemented with a variety of trained neural networks means that the functional complexity of technology could see exponential increase in the near future. As far as we are concerned, the more the better.

Source: Energy-friendly chip can perform powerful artificial-intelligence tasks | MIT News | Massachusetts Institute of Technology

Replies

You are reading an archived discussion.

Related Posts

Whats the coefficient of drag for a Lamborghini car?(Rough estimate) N also the Coefficient of lift
We introduce ourself's as an education startup. We run the website mechieprojects.com We are a group of highly motivated engineers with the zeal to help current batch of engineering students...
Hi guys, I'm using an electromagnet to detect the presence of a steel door through analog pins on an Arduino. This works great if the magnet itself isn't under power,...
Hey there! I am a hobbyist and a budding electronics engineer from Mumbai and I wanted to buy an analog used oscilloscope for cheap, 2channel 20-50Mhz. I tried ebay, quikr,...
This is what in my mind.I want to make a system on chip like Raspberry pi,Beaglebone,etc..Actually,I am a second year Electronics & Telecommunication student.So,I'm actually planning about my third year...